Whilst recent years have witnessed the development and exploitation of operational Earth Observation (EO) satellite constellation data, the valorisation of historical archives has been a challenge. The European Space Agency (ESA) Landsat Multi Spectral Scanner (MSS) products cover Greenland, Iceland, Continental Europe and North Africa and expect to represent an archive of over 600,000 processed Level 1 (L1) scenes to join the 1 million ESA Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) products already available. The ESA Landsat processing baseline has been significantly updated to account for MSS data ageing and to accommodate for remote sensing techniques used during the past. ESA began acquiring MSS data in 1975 and it is known that the old data can be degraded including accuracy lost and missing data. For these reasons, the content of the product format has been reviewed and significantly updated to make products fit for user purposes. This paper presents the new MSS product format including the updated metadata parameters for error traceability and the specification of the Quality Assurance Band (BQA) engineered to allow the best pixel selection and also the application of image restoration schemes. This paper discuss also major improvements done in radiometric and geometric processing. For the benefits of the community, ESA is now able to maximize the number of L1 MSS products that can be potentially generated from the raw Level 0 (L0) data and the highest possible data quality is reached. Also, by improving product format, processing and adding a pixel based quality band, the old MSS archive becomes interoperable with recent data from on-going missions by way of assuring product quality on a pixel basis.

Restoration of Very High Resolution (VHR) optical Remote Sensing Image (RSI) is critical and leads to the problem of removing instrumental noise while keeping integrity of relevant information. Improving denoising in an image processing chain imply increasing image quality and improving performance of all following tasks operated by experts (photo-interpretation, annotation, cartography) or by algorithms (land cover, land use, change detection, 3D reconstruction). Instrumental noise is known and follows a Poisson distribution which is converted to additional Gaussian noise with Anscombe transform. In a context of large industrial VHR image production, the selected denoising method should optimized 3 objectives: (1) accuracy and robustness with relevant information and saliency conservation, (2) rapidity due to huge amount of data acquired and/or archived and (3) flexibility with adaptation to image content. Very recent research in image processing leads to a fast and accurate algorithm called Non Local Bayes (NLB) [Buades, 2013] that we propose to adapt, optimize and improve for VHR RSI. This method is well suited for mass production thanks to its best tradeoff between accuracy and computation complexity compared to other state-of-the-art methods. NLB is based on a simple principle: similar structures in image have similar noise distribution and thus can be denoised with the same noise estimation. Non Local Bayes operates as follow: for each pixel of the image, (1) similar patches are found around the pixel and (2) a Bayesian estimation of the denoised pixel is computed. In this article, we will describe in details algorithm operations and performances. We will also propose improvements and optimizations to decrease computation complexity while keeping high denoising accuracy. Finally, we will show and explain denoising results on various typical real areas: land, water; urban, airport, etc.

In the frame of the Copernicus programme, ESA has developed and launched the Sentinel-2 optical imaging mission that delivers optical data products designed to feed downstream services mainly related to land monitoring, emergency management and security. The Sentinel-2 mission is the constellation of two polar orbiting satellites Sentinel-2A and Sentinel-2B, each one equipped with an optical imaging sensor MSI (Multi-Spectral Instrument). Sentinel-2A was launched on June 23rd, 2015 and Sentinel-2B followed on March 7th, 2017. With the beginning of the operational phase the constellation of both satellites will enable image acquisition over the same area every 5 days or less. To use unique potential of the Sentinel-2 data for land applications and ensure the highest quality of scientific exploitation, accurate correction of satellite images for atmospheric effects is required. Therefore the atmospheric correction processor Sen2Cor was developed by Telespazio VEGA Deutschland GmbH on behalf of ESA. Sen2Cor is a Level-2A processor which main purpose is to correct single-date Sentinel-2 Level-1C Top-Of-Atmosphere (TOA) products from the effects of the atmosphere in order to deliver a Level-2A Bottom-Of-Atmosphere (BOA) reflectance product. Additional outputs are an Aerosol Optical Thickness (AOT) map, a Water Vapour (WV) map and a Scene Classification (SC) map with Quality Indicators for cloud and snow probabilities. Telespazio France and DLR have teamed up in order to provide the calibration and validation of the Sen2Cor processor. This presentation provides an overview over the Sentinel-2 data, processor and products. It presents some processing examples of Sen2Cor applied to Sentinel-2 data, provides up-to-date information about the Sen2Cor release status and recent validation results at the time of the SPIE Remote Sensing 2017.

In MS pansharpening, the injection model ruling the combination of MS with the spatial detail extracted from P is stated between the resampled MS bands and lowpass-filtered P, PL, having same spatial frequencies as MS. Popular models are: i) the projection model, representing the basis of the Gram-Schmidt spectral sharpening and of the context-based decision (CBD); ii) the multiplicative or modulation model, which is the basis of HPM, Brovey, SVR, UNB pansharp, SFIM and AWLP. SFIM gave an interpretation in terms of radiative transfer model: the MS bands interpolated at the scale of P represent a low spatial resolution spectral radiance; P represents the high spatial resolution solar irradiance. Thus, PL is a low-resolution irradiance and conveys radiometric information at the spatial resolution of MS. A high resolution MS (HRMS) is synthesized at each pixel as low resolution MS (LRMS) divided by PL and multiplied by P. Since the ratio of spectral radiance to solar irradiance is an estimate of spectral reflectance, when LRMS/PL is multiplied by P, a low-resolution reflectance is multiplied by a high resolution irradiance. So far, few authors have considered the path radiance, which is an energy scattered by the atmosphere that reaches the aperture of the instrument without being reflected by the Earth’s surface. The path radiance of each band, which appears as a haze in a color composite display, should be estimated and removed before the modulation is accomplished and re-inserted after the sharpening. In this paper, several methods for estimating the path radiance, one empirical, another statistical and a third based on modelling the atmosphere (Fu-Liou-Gu), are compared in terms of benefits on the quality of fusion, both CS and MRA. Experiments on true MS images show that the advantages are consistent in terms of spectral quality and especially significant on vegetation.

Recently spatial resolution enhancement of hyperspectral images (HSI) has recently gained particular attention in many practical applications. Some of popular superresolution techniques are based on fusion of HSI with high spatial resolution image such as multispectral images (MSI). Also due to the limitation of sensors, Gaussian and non-Gaussian noises affect the HSI quality. These noises have reverse effects on superresolution results. In this paper, a combination method for denoising and fusion of HSI with MSI is introduced based on Shearlet transform (ST) and Bayesian Sparse (BS) representation. ST reduces the Gaussian and Non-Gaussian noises and BS increases the spatial resolution of HSI. The proposed algorithm is applied to the real HSI datasets. The obtained results of proposed method in comparison with some of state-of-the-art fusion algorithms based on spectral unmixing and sparse representation methods especially on noisy bands show that the proposed method significantly increases the spatial resolution of HSI and decreases the spectral distortion efficiently.

In this paper, the application of super resolution (SR, restoring a high spatial resolution image from a series of low resolution images of the same scene) techniques to GaoFen(GF)-4, which is the most advanced geostationary-orbit earth observing satellite in China, remote sensing images is investigated and tested. SR has been a hot research area for decades, but one of the barriers of applying SR in remote sensing community is the time slot between those low resolution (LR) images acquisition. In general, the longer the time slot, the less reliable the reconstruction. GF-4 has the unique advantage of capturing a sequence of LR of the same region in minutes, i.e. working as a staring camera from the point view of SR. This is the first experiment of applying super resolution to a sequence of low resolution images captured by GF-4 within a short time period. In this paper, we use Maximum a Posteriori (MAP) to solve the ill- conditioned problem of SR. Both the wavelet transform and the curvelet transform are used to setup a sparse prior for remote sensing images. By combining several images of both the BeiJing and DunHuang regions captured by GF-4 our method can improve spatial resolution both visually and numerically. Experimental tests show that lots of detail cannot be observed in the captured LR images, but can be seen in the super resolved high resolution (HR) images. To help the evaluation, Google Earth image can also be referenced. Moreover, our experimental tests also show that the higher the temporal resolution, the better the HR images can be resolved. The study illustrates that the application for SR to geostationary-orbit based earth observation data is very feasible and worthwhile, and it holds the potential application for all other geostationary-orbit based earth observing systems.

Many researchers have developed algorithm for generating the super resolved maps from the remotely sensed satellite images with heterogeneous class components. Each algorithm has its own advantages and limitations. Slow rate of convergence, low accuracy, local optimum solution are some of the issues in the super resolution mapping algorithms. To overcome the problems, a new soft classification algorithm (New Fuzzy Cluster Centroid for Unsupervised Classification Algorithm) and super resolution mapping algorithm (Multi-objective Genetic Algorithm) was developed and demonstrated in this paper. Super Resolution Mapping consists of two steps. The first step is the generation of proportion images from the satellite image using the soft classification algorithm, New Fuzzy Cluster Centroid (NFCC) for Unsupervised Classification Algorithm. The proportion images contains the proportion or percentage of class components within a pixel. The second step is the generation of super resolved map from the proportion images using super resolution mapping algorithm, multi-objective genetic algorithm. In NFCC, a new objective function is formulated which increases the stability of the algorithm and reduces the number of iterations for image classification. In genetic algorithm, multiple objectives and the concept of elitism used which improves the performance of the super resolution mapping. Experiments are carried out with the use of multi-date Landsat OLI image of the drinking water reservoir, Puzhal of Chennai City. The super resolved map is generated for the water spread area of the reservoir. From the super resolved map, water spread area of the reservoir is accurately estimated. The water-spread area estimated using super resolution mapping approach can be used as an input to estimate the volume at different water levels of the reservoirs. This super resolution mapping approach gives minimum error when compared to other sub-pixel, per-pixel approaches.

Sentinel-2 Global Land Cover project (S2GLC) aims at developing procedures and providing recommendations for deriving land cover classifications on a global scale from the imagery of Sentinle-2 constellation. By assuring high temporal resolution the Sentinel-2 dataset allows revision of existing and development of new solutions for multitemporal image analysis. The study described here is a part of the S2GLC research project that presents a method for combining results of land cover classification performed on individual Sentinel-2 images acquired at different time periods. The general aim of this methods is to derive a map that represents the most reliable land cover information. The study starts with a pre-processing part consisting of an atmospheric correction and a cloud mask derivation as obligatory procedures for a multitemporal data analysis. This is followed by supervised classification of Sentinel-2 images with Boosting, Random Forest and Support Vector Machine algorithms and classification features (i.e. spectral, statistical, structural and multitemporal). The number and acquisition dates of the classified images are based on the user selection and should reflect different conditions met in the landscape and assure better differentiation of classes when multitemporal observations are combined. However, the final selection is also strongly dependent on the amount and distribution of cloud-covered areas within an individual satellite scene. The goal of this analysis, which is under development, is to elaborate the most efficient and reliable aggregation method. Important aspect of the aggregation relates to the method of assigning a class label to a given pixel in the resulting land cover map. In our work we examine such factors as the number of occurrence of a certain land cover class within a time series, the level of fitting of a classified pixel to the training data of a given class (prediction score resulting from a classifier) and a combination of both factors. Due to the global approach of the project our methods are being tested on study areas located in various parts of the globe representing different landscape and climate conditions and characterised by various combination of land cover types and degree of cloudiness. The initial results show that the mapping accuracy resulting from our aggregation method outperform accuracies of individual classifications within the time series. They also show that besides the class aggregation strategy itself the availability of cloud free images and the quality of a cloud mask and alignment of images from a time series are critical.

Exploitation of temporal series of hyperspectral images is a relatively new discipline that has a wide variety of possible applications in fields like remote sensing, area surveillance, defense and security, search and rescue and so on. In this work, we discuss how images taken at two different times can be processed to detect changes caused by insertion, deletion or displacement of small objects in the monitored scene. This problem is known in the literature as anomalous change detection (ACD) and it can be viewed as the extension, to the multitemporal case, of the well-known anomaly detection problem in a single image. In fact, in both cases, the hyperspectral images are processed blindly in an unsupervised manner and without a-priori knowledge about the target spectrum. We introduce the ACD problem using an approach based on the statistical decision theory and we derive a common framework including different ACD approaches. Particularly, we clearly define the observation space, the data statistical distribution conditioned to the two competing hypotheses and the procedure followed to come with the solution. The proposed overview places emphasis on techniques based on the multivariate Gaussian model that allows a formal presentation of the ACD problem and the rigorous derivation of the possible solutions in a way that is both mathematically more tractable and easier to interpret. We also discuss practical problems related to the application of the detectors in the real world and present affordable solutions. Namely, we describe the ACD processing chain including the strategies that are commonly adopted to compensate pervasive radiometric changes, caused by the different illumination/atmospheric conditions, and to mitigate the residual geometric image co-registration errors. Results obtained on real freely available data are discussed in order to test and compare the methods within the proposed general framework.

Small family farms dominate agricultural production in the developing world, producing 80% of the world’s food. More than 500 million family farms manage between 70 and 80% of farmland, particularly in sub-Saharan Africa, South America, and Asia. Understanding how weather and climate affect growing conditions for these smallholders is a central part of food security assessment. Due to the diversity of farm family’s access to resources and sources of income, there is no direct connection between changes in food production and food security outcomes. However, human health beyond nutrition is likely to be affected by environmental changes occurring across multiple geographic and time scales. Impacts range from increasing transmissibility and the range of vectorborne diseases, such as malaria and yellow fever, to undermining nutrition through deleterious impacts on clean water, changes in access to diverse environmental resources and concomitant increases in food prices. Using a variety of analytic approaches, this talk will describe how remote sensing data and weather information can be combined with household survey and consumption information to better understand the impact of environmental and climate change on food security outcomes. New approaches from the fields of information and communication technology (ICT) and big data analytics will also be described, as they are enabling new tools that will help localize agricultural information and connect farmers with appropriate expertise. Using mobile technologies to gather specific information on small farmers, including crop grown, area cultivated, yield, access to resources, use of inputs and labour, we can improve our ability to generate much higher quality agricultural statistics, which can further improve assessment of food production outcomes. Remote sensing data can be much more useful to smallholder family farmers if they can be integrated into their decision making. Small farmers make up a large proportion of the food insecure populations, therefore promoting their increased productivity through information from remote sensing is a key way to spur economic growth, environmental resilience and improved food security.

Hyperspectral imagery has significantly contributed to the progress of remote sensing studies. The hyperspectral imagery is formed by a several narrow and continuous bands covering different regions of the electromagnetic spectrum, such as spectral bands of the visible, near infrared and far infrared. Hyperspectral imagery provides extremely higher spectral resolution than high spatial resolution multispectral imagery, improving the detection capability of terrestrial objects. The greatest difficulty found in the hyperspectral processing is the high dimensionality of these data, which brings out the 'Hughes' phenomenon. This phenomenon specifies that the size of training set required for a given classification increases exponentially with the number of spectral bands. Therefore, the dimensionality of the hyperspectral data is an important drawback when applying traditional classification or pattern recognition approaches to this hyperspectral imagery. In our context, the dimensionality reduction is necessary to obtain accurate thematic maps of natural protected areas. Dimensionality reduction can be divided into the feature-selection algorithms and feature-extraction algorithms. We focus the study in the feature-extraction algorithms like Principal Component Analysis (PCA), Minimum Noise Fraction (MNF) and Independent Component Analysis (ICA). After a review of the state-of-art, it has been observed a lack of a comparative study on the techniques used in the hyperspectral imagery dimensionality reduction. In this context, our first objective was to perform a comparative study of the traditional techniques of dimensionality reduction (PCA, MNF and ICA) to evaluate their performance in the classification of high spatial resolution imagery of the CASI (Compact Airborne Spectrographic Imager) sensor. On the other hand, previous studies have used some bands from these dimensionality reduction techniques to obtain suitable classification maps. In this context, our second objective was to obtain the most accurate classification map using a suitable fusion of a selection of those transformed bands having more information and separability from each dimensionality reduction technique.

In this paper, we introduce a new unsupervised classifier for HSI by using image segmentation and spectral unmixing. Supervised classification algorithms such as support vector machine (SVM) require the ground truth of HSI. In these algorithms, some pixels for randomly training and remaining pixels for testing are considered. However, in some practical applications, the ground truth doesn't exist. In the proposed method we use the concept of spectral unmixing for unsupervised classification. In this case first, the number of classes is considered equal to the number of endmembers. Second, the endmember matrix is defined. Third, the abundance fraction matrices are calculated. Fourth, the initial ground truth is constructed by considering the endmember number of maximum absolute value of each abundance fraction as the class number. Fifth, each pixel which has the same eight neighboring (vertical, horizontal and diagonal) class is a good candidate for training data and after that, some of the good candidate pixels are randomly selected as final training data and remaining pixels are considered as testing data. Sixth, SVM is applied to the data. Steps five and six are iteratively repeated. In order to validate the efficiency of the proposed algorithm, two real datasets are used. The obtained classification results are compared with some of the state-of-the-art classification algorithms and the classification accuracy of the proposed method is close to supervised algorithms.

We address the problem of hyperspectral image (HSI) pixel partitioning using nearest neighbor - density-based (NN-DB) clustering methods. NN-DB methods are able to cluster data objects without specifying the number of clusters to be found. Among the NN-DB approaches, we focus on deterministic methods, e.g. ModeSeek, knnClust, and GWENN (standing for Graph WatershEd using Nearest Neighbors). These methods only require the availability of a k-nearest neighbor (kNN) graph. Recently, a new NN-DB clustering method, called Density Peak Clustering (DPC), has been proposed and received much attention from the data scientists’ community, and kNN versions of it have been proposed since and shown their efficiency. However, they still suffer from the difficulty of obtaining the kNN graph due to the quadratic complexity with respect to the number of pixels. This is why GWENN was embedded into a multiresolution (MR) scheme to bypass the computation of the full kNN graph over the image pixels. In this communication, we propose to extent the MR-GWENN scheme on three aspects. Firstly, similarly to knnClust, the original labeling rule of GWENN is modified to account for local density values, in addition to the labels of previously processed data objects. Secondly, we set up a modified NN search procedure within the MR scheme, in order to stabilize of the number of clusters found from the coarsest to the finest spatial resolution. Finally, we show that these extensions can be easily adapted to the three other NN-DB methods (ModeSeek, knnClust, knn-DPC). Experiments are conducted to compare the four NN-DB methods for pixel clustering in HSIs. We show that NN-DB methods can outperform classical clustering methods such as FCM, in terms of classification accuracy, relevance of found clusters, and clustering speed. Finally, we demonstrate the feasibility and evaluate the performances of NN-DB methods on a very large image acquired by our AISA Eagle hyperspectral imaging sensor.

In signal and image processing principal component analysis (PCA) is often used for dimensionality reduction and feature extraction in pre-processing steps to for example classification. In remote sensing image analysis PCA is often replaced by maximum autocorrelation factor (MAF) or minimum noise fraction (MNF) analysis. This is done because MAF and MNF analyses incorporate spatial information in the orthogonalization of the multivariate data which is conceptually more satisfactory and which typically gives better results. In this contribution, autocorrelation between the multivariate data and a spatially shifted version of the same data in the MAF analysis is replaced by the information theoretical, entropy and Kullback-Leibler divergence based measure mutual information. This potentially gives a more detailed decomposition of the data. Also, the orthogonality between already found components and components of higher order requested in the MAF analysis is replaced by a requirement of minimum mutual information between components. The sketched methods are used on the well-known AVIRIS (https://aviris.jpl.nasa.gov/) Indian Pines data.

In this paper a new supervised method for classification of Hyperspectral Images (HSI) based on 2D Non-Subsampled Shearlet Transform (NSST) and Minimum Noise Fraction (MNF) is introduced. Recently some methods based on Wavelet Transform for HSI classification have been introduced. But wavelets are known to successfully represent 1D signals with high sparsity. However, 2D wavelet representation is not highly sparse. In recent years, multidirectional representations based on contourlets, curvelets, and etc. have been introduced in order to efficiently handle the geometrical features of multidimensional signals. The shearlet transformation was introduced and shown that it has a superior directional selectivity which makes it an optimal sparse representation of images. In this paper a special type of discrete shearlet transform “Non-Subsampled Shearlet Transform (NSST)” is used. In this paper, HSI are considered as a 3D dataset: two spatial dimensions and one spectral dimension. NSST is applied to each spectral bands of HSI. After that, because the NSST coefficients at any level of decomposition preserve the same size of the original signal even though the number of subbands grows as more and more levels of decomposition applied. Then, MNF is applied to all NSST coefficient subbands in order to reduce the dimensionality reduction. The MNF is applied to all NSST coefficient subbands in order to reduce the dimensionality reduction. MNF sort the informative bands based on signal to noise ratio (SNR). The bands with high SNR are preserved. Finally, the Support Vector Machine (SVM) classifier is used in order to classify the image based on these extracted features. In order to validate the efficiency of the proposed algorithm, two real datasets are used. The obtained classification results are compared with some of the state-of-the-art classification algorithms and the proposed method has reached the highest classification accuracy.

The very high spectral resolution of Hyperspectral Images (HSIs) enables the identification of materials with subtle differences and the extraction subpixel information. However, the increasing of spectral resolution often implies an increasing in the noise linked with the image formation process. This degradation mechanism limits the quality of extracted information and its potential applications. Since HSIs represent natural scenes and their spectral channels are highly correlated, they are characterized by a high level of self-similarity and are well approximated by low-rank representations. These characteristic underlies the state-of-the-art in HSI denoising. However, in presence of rare pixels, the denoising performance of those methods is not optimal and, in addition, it may compromise the future detection of those pixels. To address these hurdles, we introduce RhyDe (Robust hyperspectral Denoising), a powerful HSI denoiser, which implements explicit low-rank representation, promotes self-similarity, and, by using a form of collaborative sparsity, preserves rare pixels. The denoising and detection effectiveness of the proposed robust HSI denoiser is illustrated using semi-real data.

During the past years several compressive spectral imaging techniques were developed. With these techniques an optically compressed version of the spectral datacube is captured. Consequently the information about the object and targets is captured in a lower dimensional space. A question that rises is whether the reduction of the captured space affects the target detection performance. The answer to this question depends on the compressive spectral imaging technique employed. In most compressive spectral imaging techniques, the target detection performance is deteriorated. We show that with our recently introduced technique, dubbed Compressive Sensing Miniature Ultra-Spectral Imaging (CS-MUSI), target detection with similar detection and false detection rates, compared to conventional hyperspectral cameras, is accomplished.

Efficient water management in agriculture requires an accurate estimation of evapotranspiration (ET). The estimation of ET at the land surface is fundamental in the determination of the exchanges of energy and mass among hydrosphere, atmosphere and biosphere in order to understand the state of agricultural areas. Even though local measures can be used to estimate the components of the energy balance, these values can not be extended to large areas due to the heterogeneity and complexity of land surfaces; and the dynamic nature of heat processes. Nowadays, there are available several balance energy surface models (e.g. SEBAL, METRIC) that provide a daily ET estimation (ETd) spatially and temporarily distributed for different crops, over wide areas. These models use sensed imagery gathered from aircraft and satellite platforms, in particular, they need infrared thermal spectral band to estimate sensible heat flux from the surface temperature. However, this spectral band is not available for most current operational remote sensors, which limits the precision of the estimations to the coarsest resolution (e.g., the thermal band of Landsat 7 has a resolution of 60 m), as well as the temporality to the dates of the available images. Even though the good results provided by machine learning (ML) methods in many different areas, few works have applied these approaches for forecasting distributed ETd on space and time where aforementioned information is missing. However, these methods do not exploit the land surface characteristics and the relationships among land covers producing estimation errors. In this work, we have developed and evaluated a methodology that provides spatial distributed estimation of ETd by means of CNNs.

Individual tree level inventory performed using high density multi-return airborne Light Detection and Ranging (LiDAR) systems provides both internal and external geometric detail on individual tree crowns. Among the forest parameters, the stem position and its Diameter at Breast Height (DBH) are the most relevant ones for accurate biomass, and forest growth estimation. However, methods that can accurately estimate these parameters along the vertical canopy are lacking in the state-of-the-art. Here, we propose a method to locate and model the stem by analyzing the empty volume created due to the stem within the 3D point cloud in the LiDAR point cloud of a conifer. The approximate location of stem is obtained by relying on the fact that conifer branches grow linearly outward from the stem in a direction nearly perpendicular to it. The approximate direction of the branches are obtained using a state-of the-art internal crown structure modelling technique. In the upper section of the crown, considering the high point density, the LiDAR points recorded most proximal to the stem location are likely due to laser reflections from the stem or branch-stem junctions. These points are also the farthest ones from the respective branch tips along the main branch direction, and tend to form a lattice of points that defines the stem location. Under the assumption that conifer stem can be closely modelled using a cone shape, we inner fit a cone to the lattice points. The parameters of this cone are used to accurately estimate the diameter at breast height, and height of the tree. The experiments were performed on a set of twenty conifers consisting of trees from five dominant European conifer species, for which the height and the DBH were known. The preliminary results prove the method to be accurate.

Building detection and characterization is one of the most active research areas in the remote sensing community. 3D-building reconstruction or urban planning are examples of important applications that benefit of accuracy building footprint estimation. In this paper, we have carried out an in-depth analysis of accuracy in the estimation of the footprint of buildings from LiDAR points. This analysis focuses on the processing steps following the object recognition, assuming that the labelling of building points have been previously performed. The most popular methods to extract the outline of objects from irregularly distributed data points have been considered. These methods are based on the extraction of the convex and concave hull, the exterior outline, and the Voronoi surround. Using the mentioned methods, the outline of a building is extracted and, then, smoothed to eliminate small angles in order to reduce the outline distortion. Finally, the Gauss method is used to compute the object area. The comparison of the approaches to estimate the outline of buildings is based on the accuracy measured from the symmetric difference between the automatic boundary estimation and the corresponding ground truth. This ground truth consists of a dataset from the OpenTopography database, containing eight buildings with different shape, size, and point density, and whose boundaries have been manually estimated. The point density has been also reduced in a half from the original point count recursively, to produce six different versions of the same building with scaled point density. The experimental results show that the error sources of the building area estimation are mainly the building shape (quantified from the corresponding circularity), the point density, and the number of points of the buildings. Regardless of these factors, the concave hull-based method ended up being the most accurate approach.

Land cover classification is one of the major research endeavors in the field of remote sensing since it provides a basis for documenting changes over time. So far, numerous classification methods have been developed. Deep learning has recently gained popularity in computer science and other disciplines. Deep learning can abstract features from numerous training data at a high level based on a series of combinations of non-linear transformation. In this study, we focused on convolutional neural networks. Since convolutional neural networks contain feature-extraction steps unlike other classification algorithms, it does not require any pre-processing steps from input images. Several studies have used deep convolutional neural networks for land cover classification, but all of them were focused on extracting spatial features. In this study, we aimed to use deep convolutional neural networks for land cover classification through the convolution of spectral curve images. We evaluated our approach using Landsat 8 time series data collected over two study sites, Gunsan-si, South Korea and Las Vegas, United States. Both study sites contain heterogeneous land cover composition, but they have different climatic and topographic characteristics. FLAASH in ENVI software was used for atmospheric correction of Landsat 8 images. Based on visual interpretation of high-resolution Google Earth images, we extracted training samples per land cover class. Multi-temporal spectral curve images were then extracted from Landsat 8 reflectance images. The convolutional neural networks were developed with four layers whose filter size was a 3x3 window. The approach proposed in this study was compared with the traditional machine learning algorithms—support vector machines, decision trees, and random forest. Results showed that the multi-temporal spectral curve-based approach produced better performance than the single-date spectral curve-based approach and the other classification algorithms. Factors that influenced the spectral curve-based classification results were carefully discussed.

Recently, there has been a noteworthy increment of using images registered by unmanned aerial vehicles (UAV) in different remote sensing applications; most of them are in an experimental stage. Sensors boarded on UAVs has lower operational costs and complexity than other remote sensing platforms, quicker turnaround times, as well as, higher spatial resolution. Concerning this last aspect, it has to pay attention on the limitations of classical algorithms based on pixels, when they are applied to high resolution images. It is well known that they provoke undesirable effects, as for example, the "salt-and-pepper" effect. In this sense, in this study, it is investigated the capability of an OBIA methodology developed “ad-hoc” for data registered by a Parrot Sequoia multispectral sensor board on a eBee SQ agricultural drone, for a particular task: the automatic generation of a digital terrain model in an agricultural area, starting from digital elevation model (DEM) and multispectral images (four bands, Green, Red-edge and NIR) generated from data registered by the eBee SQ UAV. The first step of the methodology is a segmentation step carried out by a super-pixel algorithm (SLIC). A set of spectral indices (NDVI, NDRE, GNDVI and GRVI) have been used to characterised the superpixels. Context and elevation information is used to merge the superpixels in a bottom up approach at the same time that objects that does not correspond to terrain are eliminated. The final product is the automatic procedure to generate Digital Terrain Model of the analysed area.

Estimates suggest that only 30% of the world’s population has documented and recognized land rights. Cadastres form a core ingredient of any land administration system. Cadastral maps comprise of the extent, ownership and sometimes, value, of land which are essential for recording and updating land records. Mapping cadastral/parcel boundaries using traditional, field surveys often proves to be time, cost and labour intensive. Alternate approaches, specifically, remote sensing based methods that could provide low-cost and up-to-date solutions for rapid creation and upgrading of cadastral maps are thus being investigated. Although, (semi)-automatic methods for detecting parcel boundaries are increasingly being explored, creating approaches that are both automated and transferable between contexts remains a challenge: the diversity of morphologies of parcel boundaries both within and across country contexts presents a major challenge. Object-based image analysis (OBIA) methods appear highly promising as these mimic the human interpretation process of identifying features. This study employs OBIA methods to detect visible parcel boundaries. The objective is to find the most suitable approach for extracting parcel boundaries using very high resolution (VHR) satellite images. Results prove the potential of OBIA for detecting visible cadastral boundaries but also shows that a balance between high completeness and correctness is hard to achieve: a low error of commission often comes with a high error of omission. Future work will focus on optimizing the results and finding the optimal trade-off in terms of detection and localization quality. Nonetheless, the results of our study in terms of (semi)-automatically extracted parcel boundaries can potentially support rapid and first-go cadastral parcel mapping in regions, where no reliable maps exist.

MODIS surface reflectance data is, among other medium spatial resolution multispectral imagery, the one that is distributed with the most complete quality ancillary data set. Such amount of quality information is essential for automatically selecting the highest quality MODIS daily images, using, for instance, a geostatistical analysis of spatial pattern images based on variogram tools. However, this automatic selection certainly could depend on the spectral information of each MODIS band. This work studies the effect of each spectral band pattern on the variogram analysis and it aims at the identification of the most suitable spectral band for the spatially characterization of a given geographical region. The analysis consists in testing the influence of each of the reflectance bands and the first component of the PCA for each of the MODIS MOD09GA product (500 m spatial resolution), corresponding to the year 2009. The study region selected, Catalonia, northeast Iberian Peninsula, covers an area of 32000 km2. Specifically, the combination of the MODIS quality data set information and the variogram analysis of reflectance data allows the detection of different types of anomalies in the variability among the pixels of a given image that are identified through its correspondence with the fitted variogram parameters: nugget, sill and range. The results are threefold: first, the variogram analysis is reaffirmed as an extremely useful technique for automatically selecting remote sensing high quality images, second, the role of the spectral bands on the spatial pattern variability of the variogram is confirmed, and third, the analysis highlights the potential of high computational effort needed for huge processing, computing the variogram analysis of a big set of images. Therefore, it is crucial to select the appropriate spectral band in order to optimize and substantially improve the automatic selection of remote sensing images using geostatistical analysis based on variogram tools.

In this paper we have used two statistics in the context of change detection in mono-polarized Synthetic Aperture Radar (SAR) imagery. The data used in this study is a multi-temporal pair of HH channel intensity RADARSAT-2 images of an agricultural scene in Winnipeg, Manitoba which was acquired on 7th and 31st of July 2012. One of the statistics is derived from a stochastic distance viz., Hellinger distance, and the other is based on the Shannon entropy which gives a measure of randomness. We have assumed the Gamma model for the HH channel intensity data with mean and number of looks as the two free parameters. Test statistics are often used to design hypothesis tests after theoretically deriving their asymptotic distributions. Again these hypothesis tests to have a practical utility, require a judicious choice of sample size, and a level of significance for thresholding. Instead of relying on the validity of the asymptotic distribution of test statistics (i.e. an implicit assumption for using p-values and level of significance), in this work we use the test statistics as direct quantifiers of change. A simple k-means clustering with k = 2 is applied on these quantifiers which segregates change and no-change regions. The two information theoretic measures chosen provide substantive evidence for change detection. The corresponding change maps are studied together to understand the complementary nature of the selected statistics. It is inferred that these two statistics may be used in tandem for better change detection analysis in SAR imagery.

Azimuth Ambiguities (AA) are one of the main sources of false alarm in ship detection and their identification plays an important role in the optimization of automatic ship detection procedures. We present the implementation of a procedure to adapt an asymmetric Wiener filtering (AWF) methodology aimed to detect and discard ghost signal due to AA in SAR images for X-band Cosmo Sky Med (CSK) images. Many methods to face with this problem were proposed and presented in recent literature. After a review of different approach to this problem, we describe the procedure to adapt the AWF to the case of X-band CSK images by implementing a selective blocks approach. The application of the described method to the Cosmo SkyMed images is not trivial because of the very poor sampling of the Azimuth Antenna Pattern (AAP) released. This leak of information required to generate a synthesized profile of the AAP. Moreover, we considered DGM_B (spectro-detected) images and not SLC/SCS as commonly found in literature. This approach revealed to be correct; preliminary results are satisfactory. Another important issue was the definition of the filtering thresholds for the binary images that are needed to select the areas where are the AAs and that must be substitute in the input image. The solution was to adopt a blocks approach so that the original image is divided in blocks (same number of columns but different rows) and the normalization and the filter calculation is repeated over each block.

Several approaches have been used in remote sensing to integrate images with different spectral and spatial resolution in order to obtain fused enhanced images. The objective of this research is three-fold. To implement in R three image fusion techniques: High Pass Filter, Principal Component Analysis and Gram-Schmidt; to apply these techniques to merge multispectral and panchromatic images from five different images with different spatial resolutions; finally, to evaluate the results using the universal image quality index (Q index) and the ERGAS index. Regarding to the qualitative analysis, Landsat-7 and Landsat-8 show greater color distortion with the three pan-sharpening methods; whereas in the other images results were better. Q index revealed that HPF fusion performs better in the QuickBird, IKONOS and Landsat-7 images, followed by the GS fusion; whereas in the case of Landsat-8 and Natmur-08 images, the results were much evener. Regarding the ERGAS spatial index, the ACP algorithm performed better for the QuickBird, IKONOS, Landsat-7 and Natmur-08 images, followed closely by the GS algorithm. Only for the Landsat-8 image, the GS fusion presented the best result. Whereas in the evaluation of spectral components, HPF results tended to be better and ACP results worse, the opposite results occurred with the spatial components. Better quantitative results are obtained in Landsat-7 and Landsat-8 images with the three fusion methods compared to the QuickBird, IKONOS and Natmur-08 images. This contrasts with the qualitative evaluation reflecting the importance of splitting the two evaluation approaches (qualitative and quantitative). Significant disagreement may arise when different methodologies are used to asses the quality of an image fusion. Moreover, it is not possible to designate, a priori, an algorithm as the better, not only because of the different characteristics of the sensors, but also because of the different atmospherics conditions or peculiarities of the different study areas, for example.

In order to objectively compare the two infrared cameras ones must measure and compare their parameters on a laboratory test stand. One of the basic parameters for the evaluation of a designed camera is NEDT. In order to examine the NEDT, parameters such as sensitivity and pixels noise must be measured. To do so, ones should register the output signal from the camera in response to the radiation of black bodies at two different temperatures. The article presets an application and measuring stand for determining the parameters of microbolometers camera. Prepared measurements were compared with the result of the measurements in the Institute of Optoelectronics, MUT on a MST test stand by CI SYSTEM. This test stand consists of IR collimator, IR standard source, rotating wheel with test patterns, a computer with a video grabber card and specialized software. The parameters of thermals cameras were measure according to norms and method described in literature.

Synthetic aperture interferometric radiometers (SAIR) are powerful instruments for high spatial resolution observation of planetary surfaces. Recently, it is being introduced to human security applications, because the brightness temperatures of human body and prohibited items are different and passive mode eliminates security concerns about the damage to human body. However, the most important problem is that the contours and details of prohibited items are blurry in rebuild image, which makes it difficult to identify prohibited items. The relationship between visibility function (V) and brightness temperature (T) is no longer fourier transformation in near-field. In this study, a regularization model based on G-matrix model (V=GT) is proposed by introducing a penalty function, which is represented as a function of gradient of the image. For keeping the contour feature, the penalty function must meet the following two rules: 1) in no contour area, the function shows the characteristic of smooth and isotropic; 2) in contour area, the function is smooth in the gradient direction and rough in gradient orthogonal direction. At last, an appropriate penalty function has been found. However, at this point, the question arises on the optimum value of the regularization parameter. The regularization parameter controls the tradeoff between the image noise and sharpening degree. Usually, the regularization parameter is small enough empirically to keep the contour in the image. Simulation for SAIR is performed to validate the contour enhancement imaging method. A complex scene consisting of many small area with different shape and brightness temperature value corresponding to different prohibited items is created. The different results are displayed with the increase of regularization parameter. And the reconstructed image by proposed method is compared with the results by inverse fourier transformation method, near-field focusing algorithm and moore-penrose method. The proposed method show better image quality.

This paper presents a coherent inverse synthetic-aperture imaging ladar(ISAL)system to obtain high resolution images. A balanced coherent optics system in laboratory is built with binary phase coded modulation transmit waveform which is different from conventional chirp. A whole digital signal processing solution is proposed including both quality phase gradient autofocus(QPGA) algorithm and cubic phase function(CPF) algorithm. Some high-resolution well-focused ISAL images of retro-reflecting targets are shown to validate the concepts. It is shown that high resolution images can be achieved and the influences from vibrations of platform involving targets and radar can be automatically compensated by the distinctive laboratory system and digital signal process.

As the improvement of temporal, spatial, spectral resolution of the satellite sensor, remote sensing technique becomes an important way to monitor the surface in local and global scale. How to accurately calibrate the data is a primary problems. During the past several decades, there are many radiometric calibration methods have been developed. As an important post launch method, not only can be used to evaluate the on-board calibrators but also can be allowed for a traceable knowledge of the absolute accuracy although it has the drawbacks of low frequency data collections due expensive on personal and cost. To overcome the problems, CEOS Working Group on Calibration & Validation (WGCV) Infrared Visible Optical Sensors (IVOS) subgroup has proposed an Automated Radiometric Calibration Network (RadCalNet) project. Baotou site is one of the four demonstration sites of RadCalNet. The superiority characteristics of Baotou site is the combination of various natural scenes and artificial targets. The main targets for automated radiometric calibration is a three-gray-level permanent targets. The targets is distributed as a chess box. Each single target has a size of 48m×48m.The site is equipped with automated sunphotometer and meteorological station. Besides, in each target, an automated spectrum measurement instrument is developed to obtain the surface reflected radiance spectra every 2 minutes with a spectrum resolution of 2nm. To meet the requirement of RadCalNet, a surface reflectance spectrum retrieval method is used to generate the standard input files. Then the TOA reflectance spectra are derived from the input files. In order to overcome adjacent effect due to the limited size of targets, an adjacent-effect-correction method has also been proposed. Now, the whole automated radiometric calibration system has provided products nearly one year. The results of the demonstration satellites, including Landsat 8, Sentinal-2A, show that the overall accuracy can be within 5%.

Band selection (BS) is one of the most important topics. The objective of BS is to find a set of representative bands that can represent the whole image with lower inter-band redundancy. Many types of BS algorithms have been proposed in the recent years. However, most of them belong to "Off-line" methods. It means that the algorithms can only be implemented on the collected data. Those off-line methods are sometime useless since many HSI applications are timeliness, particular in disaster prevention and target detection. To tackle this issue, a new concept, called Progressive Sample Processing (PSP), was proposed recently. The core concept of PSP is that it is an "On-line" framework where the specific algorithm can process the currently collected data during the data transmission under band-interleaved-by-pixel/sample (BIS/BIP) protocol. This paper proposes a new BS method that integrates a sparse-based BS method into PSP, called PSP-BS. In PSP-BS, the BS can be carried out by updating BS result recursively pixel by pixel in the same way that a Kalman filter does for updating data information in a recursive fashion. We use the self-sparse model to select the representative bands where the orthogonal matching pursuit (OMP) is used to solve the regression problem. In order to accelerate this method, we adopt two least-square updating procedures to avoid duplicate computation when receiving new pixels. The experiments conducted on two real hyperspectral datasets show that the PSP-BS can progressively output the BS status with reasonable computing time. In addition, we design several pixel sampling methods and compare their convergence properties during BIP/BIS transmission. The experiments also show that using appropriate pixel sampling method will accelerate the algorithm and achieve higher consistency.

The FAST-SIFT corner detector and descriptor extractor combination was used to automatically georeference DIWATA-1 Spaceborne Multispectral Imager and High Precision Telescope images. Features from Accelerated Segment Test (FAST) algorithm detects corners or keypoints in an image. These robustly detected keypoints have well-defined positions. Descriptors were computed using Scale-Invariant Feature Transform (SIFT) extractor. FAST-SIFT method effectively matched HPT and SMI same-subscene images detected by the NIR, B, G, and R sensors. The method was also tested in stitching NIR images with varying subscene swept by the camera. The slave image was matched to the master image. The keypoints served as the ground control points. Given the nature of the images, adjustments of characteristic thresholds of the descriptor matcher were varied to obtain the maximum matched features between two images. Random sample consensus was used to eliminate fall-out matches and ensure accuracy of the feature points from which the transformation parameters were derived. The DIWATA-1 HPT images were also matched with other images to test the sensor-invariance of the FAST-SIFT combination. Keypoints are matched based from their descriptor vector. Nearest-neighbor matching is employed based on a metric distance between the descriptors. The metrics include Euclidean, Minkowski, and city block, among others. The closest are returned. Rough matching outputs not only the correct matches but also the faulty matches. A previous work in automatic georeferncing incorporates a geometric restriction. In this work, we applied a simplified version of the learning method. RANSAC was used to eliminate fall-out matches and ensure accuracy of the feature points. This method identifies if a point fits the transformation function and returns inlier matches. The transformation matrix was solved by Affine, Projective, and Polynomial model. Accuracy of the automatic georeferencing method were determined by calculating the RMSE of interest points, selected randomly, between the master image and transformed slave image.

Polarimetric synthetic aperture radar (pol-SAR) polarimetry obtains polarimetric scattering of interesting targets when platform takes a straight, circle or other flight track. The scattering properties of interesting targets are usually considered as invariant in azimuth. In some new SAR system, such as wide-angle SAR and circular SAR (CSAR), targets are illuminated longer and look angle changes a lot. Some targets have different physical shape in different look angle. Thus complex targets can no longer be considered as having invariant scattering properties in azimuth. In fact, angle information provided by new SAR system should be considered as useful and can help to describe scattering properties of interesting targets. With enough angle information, precise description of complex target is now possible to achieve. Directivity and isotropy are important features when angle information is available. In this paper, sub-aperture analysis and polarimetric entropy are used to measure directivity and isotropy. Targets are classified into three classes based on measurement. One is targets with strong directivity. This means targets have strong scattering only in one direction. One is isotropic targets. This means targets have one certain scattering mechanism in several or all directions. One is targets with neither strong directivity nor isotropy. This means targets have random scattering mechanisms in any directions. The effectiveness of the method is demonstrated on polarimetric CSAR(pol-CSAR) data, acquired by the Institute of Electronics airborne CSAR system at P-band

Hyperspectral Images are now used in the field of agriculture, cosmetics, and space exploring. Behind this fact, there is a result of efforts to contrive miniaturization and decreasing in costs. This paper describes low-cost and small Hyperspectral Camera (HSC) under development and a method of utilizing it. Real Time Detection System for MDA is that government agencies put those cameras in small satellites and use them for MDA (Maritime Domain Awareness). MDA is required for national defense and disaster prevention especially in Japan. We assume early detection of unidentified floating objects to find out disguised fishing ships and submarines. For the analysis, there is one of the famous anomaly detections called Reed-Xiaoli Detector (RXD). RXD is considered as the inverse operation of principal component analysis (PCA). The preliminary experiment had done with a small helicopter instead of a satellite. Consequently, all objects including a person who put on a life jacket, a ship, and a buoy are recognized. In regards to HSC, it was successful to invent a small-size and light hyperspectral camera. It is of course possible to load on a satellite. There is a suggestion that MDHU makes it possible to process the large-volume data like hyperspectral data in the satellite. For the data transmission, laser communication is used. By employing 10 or more this satellites, Japan can perform MDA for Japanese territorial sea in real time.

The key to successful spectral un-mixing is indicating number of endmembers and their corresponding spectral signatures. Nevertheless, correctly estimate the number of endmembers without a priori knowledge is a very hard task because pixels in a hyperspectral image are always contain a mixture of the several reflected spectra. Currently, Noise Whitened Harsanyi, Farrand, and Chang (NWHFC) and hyperspectral signal subspace identification by minimum error (HySime) are two well-known methods for estimating the number of endmembers. However, in practice, because NWHFC requires fixing the false-alarm probability and HySime needs estimate noise of each spectral band, these two methods may not only ignore small objects but also can’t identify endmembers. In this paper, assuming endmembers in a hyperspectral image can be modeled by convex geometry. We propose a three-stage process to estimate the number of endmembers. At the first stage, principal component (PC) is used to transform original image to low-dimensional components for speeding up algorithm execution. At second stage, successive volume maximization (SVMAX) is used to obtain vertex using convex properties. At the third stage, spectral angle mapper (SAM) is used to compute similarity measures among vertex, and minimum SAM value represents vertex separation. Repeat the second and third stages by increasing transformed component dimensions until reach a predefined criteria. The number of endmembers of the image is the vertex with maximum of the vertex separations Finally, the proposed method is applied to synthetic and real AVIRIS and HYDICE hyperspectral data sets for estimating the number of endmembers. The results demonstrate that the proposed method can be used to estimate more reasonable and precise number of endmembers than the two published methods.

In this paper, Haar's generalized wavelet functions are applied to the problem of ecological monitoring by the method of remote sensing of the Earth. We study generalized Haar wavelet series and suggest the use of Tikhonov's regularization method for investigating them for correctness. In the solution of this problem, an important role is played by classes of functions that were introduced and described in detail by I.M. Sobol for studying multidimensional quadrature formulas and it contains functions with rapidly convergent series of wavelet Haar. A theorem on the stability and uniform convergence of the regularized summation function of the generalized wavelet-Haar series of a function from this class with approximate coefficients is proved. The article also examines the problem of using orthogonal transformations in Earth remote sensing technologies for environmental monitoring. Remote sensing of the Earth allows to receive from spacecrafts information of medium, high spatial resolution and to conduct hyperspectral measurements. Spacecrafts have tens or hundreds of spectral channels. To process the images, the device of discrete orthogonal transforms, and namely, wavelet transforms, was used. The aim of the work is to apply the regularization method in one of the problems associated with remote sensing of the Earth and subsequently to process the satellite images through discrete orthogonal transformations, in particular, generalized Haar wavelet transforms. General methods of research. In this paper, Tikhonov's regularization method, the elements of mathematical analysis, the theory of discrete orthogonal transformations, and methods for decoding of satellite images are used. Scientific novelty. The task of processing of archival satellite snapshots (images), in particular, signal filtering, was investigated from the point of view of an incorrectly posed problem. The regularization parameters for discrete orthogonal transformations were determined.

The quality of archeological sites documenting is of great importance for cultural heritage preserving and investigating. The progress in developing new techniques and systems for data acquisition and processing creates a great basis for achieving a new quality of archeological sites documenting and visualization. Archeological data has some specific features which have to be taken into account when acquiring, processing and managing. First of all it is a needed to gather as full as possible information about findings providing no loss of information and no damage to artifacts. So the most powerful and convenient means which satisfy this requirement is remote sensing technologies. A novel approach to archeological data acquiring and fusion is presented. It combines a set of photogrammetric techniques for obtaining geometrical and visual information at different scales and detailing and new approach for archeological data analysis, fusion and presenting. A set of data acquisition techniques is developed for capturing different types of information. At the most common level aerial imaging by UAV-based digital camera is performed for obtaining 3D model of the whole archeological site. A geodetic support performed by standard techniques and satellite imagery are used for accurate geo-referencing. The UAV imagery is processed by original software which implements original technique for photorealistic 3D reconstruction of the site. At the next level of data acquisition 3D scanning techniques are used for generating accurate 3D models of artifacts and detailed 3D models of environment. Also a set of high resolution color images is acquired for photorealistic presentation of the 3D models. This approach allows to combine all data in unified archeological information model with the possibility of expanding with the new sets and new types of data. The proposed approach is applied for documenting of Bosporus archeological expedition of Russian State Historical Museum.

With the rapid development of remote sensing imaging technology, the micro satellite, one kind of tiny spacecraft, appears during the past few years. A good many studies contribute to dwarfing satellites for imaging purpose, including SkyBox, Shanghai Engineering Center for Microsatellites, and so forth. Generally speaking, micro satellites weigh less than 100 kilograms, even less than 50 kilograms, which are slightly larger or smaller than the common miniature refrigerators. However, the optical system design is hard to be perfect due to the satellite room and weight limitation. In most cases, the unprocessed data captured by the imager on the microsatellite cannot meet the application need. Spatial resolution is the key problem. As for remote sensing applications, the higher spatial resolution of images we gain, the wider fields we can apply them. Consequently, how to utilize super resolution (SR) and image fusion to enhance the quality of imagery deserves studying. With an ideal sensor, we would have high-resolution color images. Unfortunately, due to spatial decimation, we only have sequent multi-frame multispectral (MS) images and panchromatic images without ideal spatial resolution. SR reconstruction is a capable way to enhance the spatial resolution of panchromatic data, whilst image sharpening is a feasible way to enhance MS imagery’s quality. Our team, the Key Laboratory of Computational Optical Imaging Technology, Academy Opto-Electronics, is devoted to designing high-performance microsat-borne imagers and high-efficiency image processing algorithms. Recently, Shanghai Engineering Centre for Microsatellites launched an experimental science microsatellite. The carried imager, designed by our team, could provide MS data and panchromatic data simultaneously. This paper proposed a strategy to realize microsat-borne MS image’s enhancement without high-resolution panchromatic image. The only tested data are a serious of same-resolution panchromatic images and MS images. We combined SR technique and a hybrid image sharpening algorithm to achieve the results. The experiments designed for the microsat-borne sensors, illustrate that the proposed strategy is feasible and performs well under real on-orbit microsatellite platform. On the aeronautic or astronautic platform, we can easily obtain the multi-frame MS image and the corresponding panchromatic image with the same resolution. Therefore, the proposed strategy can be widely used in remote sensing field.

The theory of compressed sensing(CS) has been proposed for almost a decade, and massive experiments show that CS has favorable performance in data compression and recovery, so we apply CS theory to remote sensing images. But in CS theory, the construction of classical sensing matrix for all sparse signals has to satisfy the Restricted Isometry Property (RIP) strictly, which limits applying CS in practical in image compression. While for remote sensing images, we know some inherent characteristics such as non-negative, smoothness and etc.. Therefore, the goal of this paper is to present a novel measurement matrix that breaks RIP. The new sensing matrix consists of two parts: the standard Nyquist sampling matrix for thumbnails and the conventional CS sampling matrix. A remote sensing image can be seen as a combination of sub-images which can be sparsely approximated in different frames as wavelets, curvelets and others. In this paper, we explore the sparse representation of sub-images in different domains. Since most of sun-synchronous based satellites fly around the earth 90 minutes and the revisit cycle is also short, lots of previously captured remote sensing images of the same place are available in advance. This drives us to reconstruct remote sensing images through a deep learning approach with those measurements from the new framework. Therefore, we propose a novel deep convolutional neural network (CNN) architecture which takes in undersampsing measurements as input and outputs an intermediate reconstruction image. It is well known that the training procedure to the network costs long time, luckily, the training step can be done only once, which makes the approach attractive for a host of sparse recovery problems.

Recent advances in satellite technology has led to an increased volume of remote sensing (RS) image archives, from which retrieving useful information is challenging. Therefore, one important research area in remote sensing (RS) is the content-based retrieval of RS images (CBIR). Using supervised classification methods in the context of CBIR by training the classifier with the already annotated images has attracted attention in RS. However, existing supervised CBIR systems in the RS literature assume that each training image is categorized by only a single label that is associated to the most significant content of the image. However, RS images usually have complex content, i.e., there are usually several regions within each image related to multiple land-cover classes. To overcome this limitation and to effectively characterize the high-level semantic content of RS images in supervised CBIR problems, we investigate effectiveness of different deep learning architectures in the framework of multi-label remote sensing image retrieval. This is achieved based on a two-steps strategy. In the first step, a Convolutional Neural Network (CNN) pre-trained for image classification with the ImageNet dataset is used off-the- shelf as a feature extractor. In particular, three popular architectures are explored: 1) VGG16; 2) Inception V3; and 3) ResNet50. In the second step of our research, we modify these three off-the- shelf models by fine-tuning their parameters with a subset of RS images and their multi-label information. Experiments carried out on an archive of aerial images show that fine-tuning CNN architectures with annotated images with multi-labels significantly improve the retrieval accuracy with respect to the standard CBIR methods. We find that fine-tuning using with a multi-class approach achieves better results than considering each label as an independent class.

The detection of ships from remote sensing data has become an essential task for maritime security. The variety of application scenarios includes piracy, illegal fishery, ocean dumping and ships carrying refugees. An automatic ship detection algorithm for optical pushbroom data has been developed. It makes use of the special detector assembly of most of those scanners, which allows apart from the detection of a ship also the calculation of its heading out of a single acquisition. Pushbroom scanners usually consist of more than one CCD detector line. Their mounting geometry causes a small time shift in the different channels acquired by different CCD lines. During the process of coregistration, this time shift is mostly eliminated, yielding superimposed images with static areas and objects being congruent. But, if an object was moving during acquisition, its position within channels corresponding to different CCD lines will be displaced. The Jenaoptronik camera JSS-56 carried onboard of the RapidEye satellites comprises 5 CCD lines, one for each spectral channel. The maximal time shift of about 3 seconds exists between the blue and red channel. The effect of the time shift on rapidly moving objects is striking. In a RapidEye RGB composite, one and the same airplane appears at three different positions, colored blue, green and red respectively. Since ships move considerably slower, the effect is reduced, but still clearly visible due to a significant cluster of red pixels in the direction of movement. The proposed algorithm for the automatic detection of moving ships was developed with RapidEye imagery. It consists mainly of three steps: the creation of a land-watermask, the object extraction and the examination of each single object. The latter step is build up by several spectral and geometric filters, making heavy use of the displacement caused by multiple CCD lines, finally yielding a set of ships and their direction of movement.

Wide area motion imagery (WAMI) acquired by airborne multicamera sensors enables continuous monitoring of large urban area. Each image can cover regions of several square kilometers and contain thousands of vehicles. Reliable vehicle tracking in this imagery is an important prerequisite for surveillance tasks, but remains challenging due to low frame rate and small object size. Most WAMI tracking approaches rely on moving object detections generated by frame differencing or background subtraction. These detection methods fail when objects slow down or stop. Recent approaches for persistent tracking compensate for missing motion detections by combining a detection-based tracker with a second tracker based on appearance or local context. In order to avoid the additional complexity introduced by combining two trackers, we propose an alternative single tracker framework that is based on multiple hypothesis tracking and recovers missing motion detections with a classifier-based detector. We integrate an appearance-based similarity measure, merge handling, vehicle-collision tests, and clutter handling to adapt the approach to the specific context of WAMI tracking. We evaluate the tracking framework on a region of interest of the publicly available WPAFB 2009 dataset for quantitative evaluation; a comparison to other persistent WAMI trackers demonstrates state of the art performance of the proposed approach. Furthermore, we analyze in detail the impact of different object detection methods and detector settings on the quality of the output tracking results. For this purpose, we choose four different motion-based detection methods that vary in detection performance and computation time to generate the input detections. As detector parameters can be adjusted to achieve different precision and recall performance, we combine each detection method with three different representative detector settings that yield (1) high precision and low recall, (2) high recall and low precision, and (3) best f-score. Comparing the tracking performance achieved for the different configurations allows us to quantify the sensitivity of the tracker to different types of detector errors and to derive recommendations for detector and parameter choice.

Recently, the resolution of the satellite images is getting finer and finer. In 2009, WorldView-2 achieved the 50cm at panchromatic mode, and WorldView-4, launched in 2016, supports the resolution of 31cm. Such high-resolution satellites are recognized as the effective tools for monitoring the terrestrial objects especially in surveillance and defense. For these purposes, extracting target objects from the satellite images has been one of the most important tasks of IMINT. However, the interpretation of high-resolution satellite images has been so difficult that skilled interpreters must have checked the satellite images manually because of the following issues. One is the requirement of the high detection accuracy rate. The other is the variety of the target, i.e. there are many kinds of ships, such as boat, cruise ship, cargo ship, aircraft carrier, and so on. Furthermore, there are similar appearance objects throughout the image; therefore, it is difficult even for the skilled interpreters to distinguish what object the pixels really compose. In this paper, we explore the feasibility of object extraction leveraging deep learning with high- resolution satellite images, especially focusing on ship detection. We calculated the detection accuracy using the WorldView-2 images. First, we collected the training images labelled as “ship” and “not ship”. After preparing the training data, we defined the deep neural network models to judge whether ships are existing or not, and trained them with about 50,000 training images for each label. Subsequently, we scanned the evaluation image and extracted the images predicted as “ship” by the neural network. Finally, we identified the location of each ship using the other deep neural network for segmentation. Experimental result shows the detection accuracy is over 90%, which will proves the effectiveness of the deep learning based object detection.

During the last decade, a huge number of earth observation satellites have been launched, leading to massive remote sensing (RS) data archives. Such archives require scalable and accurate methods for the quick access to RS data. In view of that, content based image retrieval (CBIR) has recently attracted great attention in the RS community [1]. To accomplish an accurate and scalable image search and retrieval in large RS data archives, hashing-based CBIR methods have been recently introduced in RS [1]. Hashing methods map high-dimensional image features into compact binary hash codes based on hash functions that can be defined by using either unlabeled images (i.e., unsupervised hashing) or a set of annotated training images (i.e., supervised hashing) [1]. Generally, supervised hashing methods achieve higher retrieval accuracy than unsupervised methods, since they provide more distinctive hash functions (and thus more distinctive hash codes). These methods describe each image by a single hash code attained by applying hash functions to global image descriptors. Moreover, they do not model possible primitives (such as different land-cover classes) present in the images. This is an important limitation as RS images generally contain several regions that may be associated to different land-cover classes and represent complex semantic content. Ignoring this in hashing-based CBIR may lead to inaccurate retrieval results, especially under complex RS image retrieval tasks. In this paper, we focus our attention on supervised hashing in RS CBIR problems. To overcome the limitations of the literature supervised hashing methods, we introduce a novel technique that aims at accurately modeling primitives (different land-cover classes) in the definition of the hashing functions. To this end, the proposed technique is defined based on a three-steps algorithm: 1) characterization of each image with descriptors of primitive classes; 2) transformation of descriptors into primitive class-sensitive multi-hash codes; and 3) assessment of the similarities between the multi-hash codes of the query image and those of each image in the archive to retrieve images similar to the query image. Experiments carried out on an archive of aerial images show that the proposed primitive class sensitive hashing method significantly improves the retrieval accuracy with respect to the standard hashing methods while keeping the same retrieval time.

We present a filtering framework based on a new statistical model for Single-Look complex SAR data – the SNIG (Scaled Normal-Inverse Gaussian). The model is based on the assumption of a geometric Brownian motion with drift. The associated compound random variable is expressed as X=βZ+√Z N, where N is a zero-mean Gaussian with variance σ2 (a measure of diffuse scattering), β is the drift (a measure of coherent scattering), and the latent variable Z is an Inverse Gaussian with unit mean and shape parameter δ (a measure of texture). The model is applied to the real and imaginary parts of complex SAR data, separately. Parameter estimation is performed using an Expectation-Maximization scheme adapted to compound probability distributions. Furthermore, the real and imaginary parts of a single SAR image are considered as mixtures of SNIGs. Separation of the mixture components (classes) is conducted using a Stochastic Expectation-Maximization algorithm. After convergence, all the pixels are clusterered into the final classes using a MAP decision rule. Model parameters are associated to each pixel according to its class, thus producing parametric images of the entire scene. A closed-form MAP filter is designed to deliver an estimate of the texture Z and uses as input the parametric images together with the original Single-Look intensity values. The result is a despeckled image, with preserved texture details. The framework is tested on RADARSAT-2 data, HV polarization, representing an image of icebergs surrounded by ocean water off the coast of the Hopen Island (Svalbard archipelago). Post-processing, the iceberg contrast-to-noise ratio (CNR) is improved compared to Single-Look Intensity images, increasing from a value of 11 to a value of 30, and the iceberg signal-to-noise ratio (SNR) is doubled compared to the same reference. We conclude that the detectability of the icebergs versus the ocean background is improved, thus demonstrating the potential of this method for iceberg detection.

Coastline detection in synthetic aperture radar (SAR) images is crucial in many application fields, from coastal erosion monitoring to navigation, from damage assessment to security planning for port facilities. First solutions have been proposed starting from edge-based segmentation [Lee and Jurkevich IEEE TGARS, 1990], also within a two-step procedure [Mason and Davenport, IEEE TGARS 1996]. More recent solutions have adopted level set segmentation either in a hierarchical scheme, also for multi-polarization images [Liu et al., IEEE JSTARS 2016], or by following a semi-supervised approach [Silveira and Heleno, IEEE GRSL 2009]. Other solutions need more than one SAR image, specifically a stack of at least three images, to extract the coastline [Baselice and Ferraioli, IEEE GRSL 2013]. This paper presents an unsupervised, computationally efficient solution to extract the coastline acquired by only one single-polarization 1-look SAR image. The method processes the original Spotlight CSK image at 1m resolution (0.5m pixel spacing) first at reduced spatial resolution (4m), and then at full spatial resolution. A preliminary segmentation is performed by jointly clustering the local variation coefficient and a radiometrically transformed version of the amplitude image. Prior knowledge for shape analysis of the resulting sea-candidate regions and accurate correction of possible focalization errors produce an accurate sea segmentation at reduced spatial resolution. Low-amplitude land regions which may be very close to the coastline are then accurately detected and processed to avoid possible errors. The final coastline extraction is performed in a computational efficient way by resorting to the original spatial resolution of the SAR image. Extensive tests on Spotlight COSMO-SkyMed images of complex coastal environments and objective assessment will demonstrate the validity of the proposed procedure which is compared to state-of-the-art methods through visual results and with an objective evaluation of the distance between the detected and the true, manually selected, coastline.

This paper presents an interferometric synthetic aperture radar (InSAR) imaging method based on L1 regularization recovered SAR images via complex approximated message passing (CAMP). The contributions of the presented method are as follows. Compared to conventional matched filtering (MF) results, the recovered images of the presented method have higher phase difference accuracy. While the presented method can reduce the computational cost efficiently by decoupling the azimuth-range couple compared with conventional L1 regularization technique. Since the presented CAMP based method can suppress the sidelobes and clutter efficiently in the recovered results which will be used to the InSAR image formation, the achieved interferogram has higher phase difference accuracy compared with MF obtained one. As typical L1 regularization recovery algorithms, e.g., iterative thresholding algorithm (ITA), could not obtain accurate phase difference, thus some applications such as InSAR processing could not be used directly via their recovered images. As an iterative recovery algorithm for L1 regularization, CAMP can not only obtain the sparse estimation of considered scene as other regularization recovery algorithms, but also a non-sparse solution with preserved background information, thus can be used to InSAR processing, which show better performance in height accuracy estimation. In sparse signal processing based SAR imaging, due to the couple of range and azimuth in raw data, conventional L1 regularization method needs to transfer the two-dimensional (2-D) echo data matrix into a vector to reconstruct the surveillance region, which produces significant computational cost. While the presented azimuth-range decouple based L1 regularization SAR imaging method solve this problem well by introducing an echo simulator to decouple the azimuth-range couple. The presented method makes large-scale sparse reconstruction of surveillance region become possible. Experiments carried out on simulated and real data confirmed the validity of the CAMP recovered images being used for InSAR processing with higher height estimation accuracy.

The presence of Antarctic subglacial lakes increases the ice flow speed and facilitates the sliding of the ice masses towards the borders of the continent, consequently rising the sea level. This has motivated the growing interest in the detection and classification of lakes at the base of the ice sheet. Indeed, over the past half century, several types of instruments and techniques allowed the identification of hundreds of subglacial lakes under the Antarctic ice sheet, as reported in the most recent subglacial lake inventory (SLI). Among such instruments, radar sounders (RS) have shown a great potential for the large scale analysis of the subglacial environment. However, the traditional manual approaches to the analysis of RS data, also called radargrams, limit the objective and effective detection and classification of subglacial lakes. Moreover, the manual analysis does not allow the joint and effective use of multiple features that can be associated to the presence of lakes. This becomes critical when considering the expected growth in RS data from both airborne and potentially future spaceborne RS missions for Earth observation. In this paper, we propose an automatic technique for the probabilistic detection of subglacial lakes in RS data. The aim of the proposed technique is to provide a soft output for the lake detection, i.e., an estimate of the probability of subglacial lake detection. This is in contrast to the available manual and automatic techniques which provide a hard decision in the classification or detection of lakes. To this aim, we propose a technique made up of two main steps, i.e., i) extraction of multiple features for the characterization of the basal interface; and ii) automatic probabilistic detection of lakes via multivariate classification. In the first step of the technique, we extract two types of features, i.e., a) structural features, which model the geometry, radar coherency and strength of basal returns; and b) statistical features, which model the statistical behaviour of the radar signals at the basal interface. Then, in the second step of the technique, all the extracted features are normalized and given in input to an automatic classifier for the final probabilistic lake detection. We applied the proposed automatic technique to a subset of radargrams acquired by the MultiChannel Coherent Radar Depth Sounder. The analysed dataset has been acquired in 2013, after the generation of the latest SLI in 2012, over about 3000 km in the East Antarctic ice sheet. The proposed method provided high overall and average classification accuracies and, interestingly, a high probability of presence of lakes at locations which are not reported in the Antarctic SLI. Thus, the proposed technique for the probabilistic detection of subglacial lakes can be a helpful tool for the glaciologic community for better analyzing the surveying region and possibly updating the current SLI.

Indonesia is periodically affected by severe volcanic eruptions and earthquakes, which are geologically coupled to the convergence of the Australian tectonic plate beneath the Sunda Plate. Multi-temporal SAR interferometry (MTI) can be used to support studying and modelling of terrain movements. The study of tectonic phenomena requires large-scale spatial analysis that poses challenges in MTI processing. A reliable modelling needs additional information coming e.g. from geodetic data, such as those provided by GNSS networks. This work is aimed at performing an analysis of ground displacements over Indonesian sites through MTI techniques. Test sites have been selected according to the availability of archived SAR data, GNSS networks, and geological data. Two stacks of COSMO-SkyMed data, both acquired in stripmap mode between 2011 and 2015, have been selected in Banda Aceh and Yogyakarta. Geological maps of the test sites are available, and several GNSS stations from the Continuously Operating Reference Stations Indonesian network are found to be located in the areas of interest. Sentinel-1 data are also available, even though their number is quite limited. Both the SPINUA and the StaMPS MTI algorithms have been used to process the data. Integration of Java and Aceh province observations from SAR satellites with ground-based GNSS observations has been attempted, with the aim of producing a uniform product, improving on the existing, low-resolution global strain rate map derived from GNSS alone. The work describes the MTI processing adopted, and the procedure developed to integrate the MTI deformation maps with the records derived from GNSS observations where possible. The ground deformations undergoing on the area of interest are then modelled according to the available products and information. Work supported by ESA project “Integrating SAR interferometry and GNSS for studying tectonic processes in Indonesia” (contract 4000114611/15/F/MOS), Alcantara Study 14-P16 “Alcantara Study Enhanced Tectonic Characterization for Indonesia”.

Several spaceborne X-band synthetic aperture radars (SARs) are nowadays operational, representing a well-established tool for Earth remote sensing at very high spatial resolution (order of meters). Until now, the exploitation of SAR for hydrological cycle modelling and numerical weather forecast has not been pursued. However, there are scientific and experimental evidences that at X band and beyond: i) atmospheric precipitation in liquid and ice phase can affect SAR imagery and its intensity can be retrieved (e.g. Marzano et al., 2012), ii) snow areal extent and mass (water-equivalent) can be detected and estimated (e.g. Montomoli et al., 2016). Within H2020 COMPET-2-2017 call, the KydroSAT mission concept has been proposed. KydroSAT foresees a miniaturised fully-digital SAR at Ku and Ka band (KydroSAR), specifically devoted to detecting and estimating atmospheric precipitation and surface snow; its baseline includes dual-polarization capability, large variable swath (50-150 km), high orbit duty cycle (>75%) and flexible ground resolution (5-150 m). The mission concept foresees the formation of two minisatellites (bringing the swath up to 150-300 km) both carrying a KydroSAR, in along-track convoy with the SAR platforms COSMO-SkyMed and SAOCOM, allowing the observation of a same scene at L, X, Ku and Ka bands. In this work, we will simulate the SAR response at L, X, Ku and Ka bands of a same scene, using the SAR forward model described in Mori et al. (2017). Simplified scenarios and 3D scenes generated by high-resolution Cloud Resolving Models (e.g., SAM) will be considered. Numerical results will be discussed in order to exploit the synergic approach of the KydroSAT convoy and to evaluate new quantitative retrievals of rain cloud parameters (e.g. columnar water content, precipitation rate). Using cryospheric synthetic scenarios, numerical simulations will be used to show the capabilities of KydroSAT for snow areal extent and SWE monitoring.

A heavy rain occurred in Hiroshima city on 20 August 2014. 107 debris flows and 59 shallow slides were induced by the heavy rain. Rapid damage assessment after natural disasters is crucial for initiating effective emergency response actions. Synthetic Aperture Radar (SAR) has great potential due primarily to its all-weather day-and-night imaging capabilities. In this study, we examine an extraction of damaged area induced by debris flows using multi-temporal COSMO-SkyMed images. Three COSMO-SkyMed images were used acquired on 6 August, 10 August and 22 August, 2014. There were one pre-disaster pair and two co-disaster pairs. The extraction methods are interferometric coherence, intensity correlation and normalized sigma nought index (NDSI) using COSMO-SkyMed image pair. These methods have been used for the assessment of building damage caused by earthquake. The Single Look Complex data of COSMO-SkyMed were co-registered each other for calculating interferometric coherence. The interferometric coherences were calculated within several moving windows. Then, interferometric coherence images were ortho-rectified using 10 m gridded Digital Elevation Model (DEM) generated by the Geospatial Information Authority of Japan (GSI). On the other hand, the Multi Look Intensity data of COSMO-SkyMed were ortho-rectified using the DEM. Then, the intensity correlation and NDSI were calculated within several moving windows. For damage assessment, normalized difference (ND) was calculated by each interferometric coherence, intensity correlation and NDSI using pre-disaster and co-disaster image pairs. The ND value increases when the land surface was changed by a disaster. Comparing with aerial photograph after the disaster and the official announcement, ND of NDSIs increased in the discharged sediment area occurred by the debris flows in the study area. In conclusion, ND of NDSIs showed good capability for extraction of damaged area induced by the debris flows at a rapid response phase.