This PDF file contains the front matter associated with SPIE Proceedings Volume 11862, including the Title Page, Copyright information and Table of Contents

Introduction to SPIE Remote Sensing conference 11862: Image and Signal Processing for Remote Sensing XXVII

Commercial smallsats are emerging as a key resource for the remote sensing community, with Planet Labs, Inc. operating the largest constellation. Global coverage is achieved through the combined efforts of many different sensors, but the unique spectral response of each sensor results in radiometric inconsistencies across the constellation. These inconsistencies have been an obstacle for researchers as the images do not match the quality the remote sensing community has come to expect from traditional platforms such as Landsat or Sentinel. Various approaches have been offered to correct cross-sensor radiometric inconsistencies in Planet image acquisitions, with many utilizing high radiometric quality platforms such as Landsat or Sentinel to normalize Planet surface reflectance via a linear transformation. These approaches have largely been applied in homogenous regions, and performance in heterogeneous landscapes is not well understood. The Planet surface reflectance images were transformed linearly using Sentinel-2 surface reflectance as a reference. Transformations were tested in six heterogeneous National Ecological Observatory Network sites where airborne hyperspectral data was available for validation. The probability density functions of the transformed Planet images and the corresponding Sentinel reference images were compared using the D statistic of the Kolmogorov-Smirnov test. The absolute spectral accuracy of the transformed Planet images was evaluated against airborne hyperspectral data. Results show that the transformations were effective in transforming the empirical cumulative distribution functions of the Planet images to be more similar to that of Sentinel regardless of initial offsets.

In geostationary Earth observation satellite system with high-resolution optical system under conceptual study phase developed by Japan Aerospace Exploration Agency (JAXA), one of its main products for users is a video with one fps captured by the telescope and the optical sensor; however, there have been few examples of satellite video with high-rate fps, especially in geostationary satellite. A satellite video has instability of the video sequences caused by an undesired disturbance of the satellite. In this study, we propose the video stabilization method corresponding to various satellite imagery characteristics, using feature point matching and bundle adjustment with eliminating outliers by using RANdom SAmple Consensus (RANSAC). Moreover, we adopt a polynomial approximation to smooth global motions of all video frames. We demonstrate the video stabilization using the experimental video captured by the helicopter. Moreover, the performance test result shows that the proposed method can stabilize the satellite video less than one pixel accuracy. In addition, we discuss various issues specific to a satellite imagery, for example, moving cloud.

The article explores the detection of unauthorized landfills based on remote sensing data. We have developed a mathematical approach that translates ordinary images to the 2-D manifold or 3-D manifolds in stereo images. At the same time, the posed task when developing algorithms is the transformation (convolution) of these manifolds into a one-dimensional sample. The developed method corresponds to the following two conditions: 1) preservation of the topological proximity of the elements of the original and expanded spaces, 2) preservation of correlations between the elements of the original and transformed spaces. The basic idea that underlies the mathematical support of the developed automated system is the concept of fractal sets. The concept of continuous orthogonal transformations, including the Fibonacci transform, is used as a mathematical basis for determining anomalous signal structures in the surrounding background. The problem of monitoring and decoding space images from the point of view of the synthesis of orthogonal systems with predetermined properties (speed of calculations, order of transformation, etc.) is presented. Examples of processing by the proposed algorithm are presented on the satellite images of the Moscow Region territories of the Russian Federation.

Image fusion methods are designed to combine multiple input images with complementary information from a scene that can increase the interpretation capabilities of the objects. The spatial or spectral quality of a fusion method highly depends on the employed fusion method and prior information exploited from input data. Spectral distortion and spatial artifacts are common drawbacks in many image fusion methods. In this study, we propose a model-based image fusion approach using sparse representation (SR) with fuzzy prior knowledge, which incorporates three constraints as prior information for better modeling of the spatial and spectral characteristics of the fusion results. These prior constraints include sparsity of coefficients, patch-wise spectral similarity, and improvement of edge details extracted from a precise unsupervised classification based on a fuzzy inference system (FIS). Edge pixels are vulnerable to spectral distortions due to the spectral mixture of neighboring pixels. Therefore, an objects-based fuzzy classification was used to separate edge pixels from background information and smooth fusion results over edge pixels to prevent spectral distortion. Moreover, in the optimization step of sparse coding, the spectral basis was estimated along with sparse coefficients in an iterative procedure. Results from the proposed fusion approach over real satellite data revealed the promising performance of the proposed SR approach based on its quantitative and qualitative results. Objective and subjective enhancement of the proposed approach compared to other well-known fusion methods included pixel-wise detail preservation and object-based spectral injection to the fused image.

In this work, we investigate on how the performance of pansharpening methods depends of their input data format, either floating-point, e.g. spectral radiance, or digital numbers (DN) in a packed fixed-point representation. It is theoretically proven and experimentally demonstrated that methods based on multiresolution analysis are unaffected by the data format, which instead is crucial for methods based on component-substitution (CS), unless the intensity component is calculated by means of a multivariate linear regression between the upsampled bands and the lowpass-filtered Pan, as it occurs for the most advanced CS methods. In an experimental setup, WorldView-2 data are either fused in their original 11-bits DN format, or converted to spectral radiance before fusion, by applying the gains and off sets provided in the file header. In the former case, fusion results are converted to spectral radiance before quality is measured. Nine fusion methods have been considered: results exactly match the theoretical investigations. For the majority of CS fusion methods, which do not feature a regression-based intensity calculation, results are better whenever they are obtained from floating-point calibrated data.

The processing of aerial high-resolution images is key for territorial mapping and change detection analysis in hydro-geomorphological high-risk areas. A new method has been developed in the context of ``CLOSE (Close to the Earth)'' project, resulting in a workflow based on open source MicMac photogrammetric suite and on High-Performance Computing. The workflow allowed to process a sequence of more than 1000 drone images captured along a reach belonging to the Basento River in Basilicata (Italy) during one single run. The workflow optimisation aims to extract the orthophotomosaic, the point cloud and the Digital Surface Model (DSM) of selected areas. The high quality of the image details can be used for land-cover classification and extrapolating features useful to mitigate the hydro-geomorphological hazard, through machine learning models trained with satellite public data. Several Convolutional Neural Networks have been tested using progressively more complex layer sequences, data augmentation and callback techniques for training procedures. The results are given in terms of model accuracy and loss.

Very high resolution satellite images can be used to generate stereoscopic digital elevation models (DEMs), efficiently and at scale, as exemplified by the upcoming CO3D mission, which aims to produce worldwide DEMs by the end of 2025. In this paper we present a deep learning stereo-vision algorithm, integrated in the Stereo Pipeline for Pushbroom Images (S2P) framework. The proposed stereo matching method applies a Siamese convolutional neural network (CNN) to construct a cost volume. A median filter is applied to every slice in the cost volume to enforce spatial smoothness, and another CNN estimates a confidence map which is used to derive the final disparity map. Simulation results on the IARPA dataset show that the proposed method improves completeness by 4.5%, compared to the state of the art. A qualitative assessment also shows that the proposed method generates DEMs with less noise.

A large number of accurate annotations of targets is a prerequisite for efficient and accurate object detection. However, to obtain such annotated samples for completing detection model training is time-consuming, laborious, and difficult to achieve. Usually, the training samples often contain noisy annotation, including mislabeled class and inaccurate bounding box. These noisy annotations reduce the classification and detection performance. In order to solve this problem and efficiently and accurately detect interested targets from remote sensing images, we propose a robust object detection method, called robustEfficientDet. In this method, firstly, with the help of EfficientDet's powerful ability of deep feature fusion, the feature representation with higher classification performance is extracted from the image. Secondly, the “Active passive losses (APL)” function is introduced into the calculation process of the classification loss to deal with the noisy annotations. In addition, in the bounding box regression, a new Focal- EIOU (short for effective Efficient Intersection over Union) loss function is introduced to reduce the positioning loss caused by inaccurate bounding box. Finally, the robustEfficientDet is constructed to improve the performance of object detection in remote sensing images. The results of several experiments show that the proposed method can achieve better detection results with noisy annotations.

Object detection in remote sensing images has far-reaching significance. However, compared with object detection tasks in the field of natural images, there are still some challenge problems that need to be improved in remote sensing, due to the similarity and unbalanced scale between objects in remote sensing images. Currently, object detection algorithms based on deep learning, i.e. Faster-RCNN, YOLO, SSD, etc., have reached an incredible grade with the concerted efforts of researchers, and have been used in various aspects such as remote sensing, face recognition, pedestrian detection and so on. However, these methods always need a lot of labeled data, while collecting a dataset is labor intensive and time consuming. In order to overcome these difficulties, we introduce a few shot learning algorithm with attention mechanism for object detection in remote sensing images, aiming to detect objects of unknown classes with only a few labeled remote sensing images. We use the structure of Siamese-network to extract the features of the target from support images and query images, and then use the features of support images as a kernel to do a depth-wise convolution on the feature map of query image. By this way, we can enhance the characteristics of the target category and weaken the characteristics of other categories and the background. For unbalanced scale problem, our method takes care of various objects of different scales in the process of extracting features. We integrate the information of multi-layer feature maps and make predictions on three feature maps of different scales. Experiments on HRSC-2016 dataset validate the effectiveness of our method.

In the last decade, the limitation of the propagation of Wildfire had become a higher necessity. In fact, it is important to optimize the resources used for dislocation to verify the probabilistic signaled fire zones. Hence, using sophisticated and low-cost techniques to sense the previously mentioned zones is highly demanded. Models with high computational necessity are not interesting for real time application. More simple models are requested, to fulfill the desired tasks with an admitted response time. Squeezesegv2 is a model applied initially for LiDAR (Light Detection And Ranging) Point Cloud data segmentation, which gives a high IoU value compared with other state of art architectures. The model was experimented in this paper, it is robust against dropout noise. Experiments were run over RGB pictures of Corsican public French dataset with 1135 RGB images. It is common that highly unbalanced datasets, which is our case, induce high precision low sensitivity. Therefore, several validation measures criterions were adopted to access the performance. In fact, the capability of the model was tested with four different metrics: Accuracy, mean Intersection over Union (IoU), Mean Boundary F1 (BF) Score, and Mean Dice coefficient. The experimental results demonstrate that the trained model, over the Corsican French dataset, with five-fold cross validation procedure can accurately detect the fire flame. The results were collected for different loss function types: Focal loss, Dice and Tversky loss. In general, the given results are very encouraging for further study using deep learning approaches.

Due to far imaging distance and relatively harsh imaging conditions, the spatial resolution of remote sensing data are relatively low. Images/videos super-resolution is of great significance to effectively improve the spatial resolution and visual effect of remote sensing data. In this paper, we propose a deep-learning-based video super-resolution method for Jilin-1 remote sensing satellite. We use explicit motion compensation method by calculating the optical flow through the optical flow estimation network and compensating the motion of the image through warp operation. After obtaining the multi-frame images after motion compensation, it is necessary to use multi-frame image fusion for super-resolution reconstruction. We performed super-resolution experiments with scale factor 4 on Jilin-1 video dataset. In order to explore suitable fusion method, we compared two kinds of image fusion methods in the super-resolution network, i.e. concatenation by channel and 3D convolution, without motion compensation. Experimental results show that 3D convolution achieves better super-resolution performance, and video super-resolution result is better than the compared single image super-resolution method. We also performed experiments with motion compensation by optical flow estimation network. Experimental results show that the difference between the image after motion compensation and the reference frame becomes smaller. This indicates that the explicit motion compensation method can compensate the difference between the frames due to the target motion to a certain extent.

One of the challenges when building Machine Learning (ML) models using satellite imagery is building sufficiently labeled data sets for training. In the past, this problem has been addressed by adapting computer vision approaches to GIS data with significant recent contributions to the field. But when trying to adapt these models to Sentinel-2 multi-spectral satellite imagery these approaches fall short. Previously, researchers used transfer learning methods trained on ImageNet and constrained the 13 channels to 3 RGB ones using existing training sets, but this severely limits the available data that can be used for complex image classification, object detection, and image segmentation tasks. To address this deficit, we present Distil, and demonstrate a specific method using our system for training models with all available Sentinel-2 channels. There currently is no publicly available rich labeled training data resource such as ImageNet for Sentinel-2 satellite imagery that covers the entire globe. Our approach using the Distil system was: a) pre-training models using unlabeled data sets and b) adapting to specific downstream tasks using a small number of annotations solicited from a user. We discuss the Distil system, an application of the system in the remote sensing domain, and a case study identifying likely locust breeding grounds in Africa from unlabeled 13-channel satellite imagery.

The task of semantic segmentation plays a vital role in the analysis of remotely sensed imagery. Currently, this task is mainly solved using supervised pre-training, where very Deep Convolutional Neural Networks (DCNNs) are trained on large annotated datasets for mostly solving a classification problem. They are useful for many visual recognition tasks but heavily depend on the amount and quality of the annotations to learn a mapping function for predicting on new data. Motivated by the plethora of data generated everyday, researchers have come up with alternatives such as Self-Supervised Learning (SSL). These methods play a deciding role in boosting the progress of deep learning without the need of expensive labeling. They entirely explore the data, find supervision signals and solve a challenge known as Pretext Task (PTT) to learn robust representations. Thereafter, the learned features are transferred to resolve the so-called Downstream Task (DST), which can represent a group of computer vision applications such as classification or object detection. The current work explores the conception of a DCNN and training strategy to jointly predict on multiple PTTs in order to learn general visual representations that could lead more accurate semantic segmentations. The first Pretext Task is Image Colorization (IC) that identifies different objects and related parts present in a grayscale image to paint those areas with the right color. The second task is Spatial Context Prediction (SCP), which captures visual similarity across images to discover the spatial configuration of patches generated out of an image. The DCNN architecture is constructed considering each particular objective of the Pretext Tasks. It is subsequently trained and its acquired knowledge is transferred into a SSL trunk network to build a Fully Convolutional Network (FCN) on top of it. The FCN with SSL trunk learns a compound of features through fine-tuning to ultimately predict the semantic segmentation. With the aim of evaluating the quality of the learned representations, the performance of the trained model will be compared with inference results of a FCN-ResNet101 architecture pre-trained on ImageNet. This comparison employs the F1-Score as quality metric. Experiments show that the method is capable of achieving general feature representations that can definitely be employed for semantic segmentation purposes.

This work addresses the problem of hyperspectral data compression and compares the reconstruction accuracy for different compression rates. Through data compression, the enormous amount of data created by hyperspectral sensors can be transmitted effectively. Remote sensing-related applications such as disaster management, land cover classification, and object recognition are improved with real-time information. We propose a 1D-Convolutional Autoencoder structure for lossy hyperspectral data compression and the necessary adjustments for realizing compression ratios of C_R = 4, C_R = 8, C_R = 16, and C_R = 32. Unlike many other contributions, we not only evaluate the reconstruction accuracy based on standard metrics like Signal to Noise Ratio and Spectral Angle Mapper but also on a target application, namely land cover classification. The reconstruction accuracy of the 1D-Convolutional Autoencoder is compared to machine learning-based lossy compression methods, namely Deep Autoencoder, Nonlinear Principal Component Analysis, and the Principal Component Analysis. The compression performances are compared using two data sets with different amounts of spectral signatures. The 1D-Convolutional Autoencoder performance surpasses the benchmark methods for all compression rates using the standard metrics. In addition, the 1D-Convolutional Autoencoder achieves the highest classification results for the land cover classification for all compression rates and is able to compress hyperspectral data efficiently. Furthermore, the robustness and generalization capability of the 1D-Convolutional Autoencoder is demonstrated by using unknown data for the evaluation.

Deep learning methods, especially convolutional neural networks(CNN), have been widely used in hyperspectral image(HSI) classification. Recently, graph convolutional networks (GCN) have shown great potential in HSI classification problem. However, the existing GCN-based methods have several problems. First, the existing methods rely too much on the adjacency matrix, which cannot be changed during training. Furthermore, most of them can only use a single kind of feature, and fail to extract the spectral-spatial information from the HSI. Finally, for the existing GCN-based methods, it is difficult to achieve the same accuracy as the mature CNN methods. In this paper, we propose a spectral-spatial hypergraph convolutional neural network (S²HCN) for HSI classification. Compared with the existing GCN-based methods, S²HCN has the following advantages. Different from the adjacency matrix that is fixed during training of GCN, S²HCN can dynamically update the weight of the hyperedge during training, which reduces the reliance on prior information to a certain extent. In addition, S²HCN generates hyperedges from the spectral and spatial features independently, and adopts the incidence matrix composed of all hyperedges to replace the adjacency matrix in GCN. In this way, the spectral and spatial features can be better integrated. Finally, compared to a simple graph structure, the hypergraph structure can express the high-dimensional relationships in the data, which is beneficial to classification problems. Sufficient experiments on two popular HSI datasets have proved the effectiveness of S²HCN.

Deep Learning (DL)-based classification schemes for hyperspectral remotely sensed data have been introduced in the last few years with remarkable success due to their capability to learn the non-linear nature of the information that conforms hyperspectral images. In particular, Convolutional Neural Networks (CNNs) have been successfully used for solving problems requiring multi-class classification in the remote sensing field involving feature extraction. CNNs operate over small cubes of the dataset called patches centered in the pixels of the image instead of relying only on the spectral information corresponding to each pixel. These networks require a high number of observations to properly produce a generalized model. In these circumstances data augmentation techniques can help alleviate the problem by generating new, synthetic samples from existing data. Imputation is a statistical technique consisting in filling or replacing missing observations or values of a subset of observations by others obtained via inference from the original dataset. In this paper, a preliminary idea for a data augmentation technique based on the use of data imputation techniques for CNN classification is presented. Different hyperspectral images of the Earth surface widely used in the remote sensing field have been considered as test datasets. The results show the viability of the preliminary idea.

n this study, one of the state-of-the-art computer vision methods, namely superpixel-based graph convolutional networks (GCN), was used to achieve an accurate semantic segmentation of SAR images. In more detail, first, simple linear iterative clustering (SLIC) is used to over-segment the input SAR image into a set of superpixels, then a feature extraction method is employed to extract features from each of the superpixels. U-Net is used as deep feature extractors. Last, GCN architecture is used for node-based classification. We thus intend to exploit the spatial informationvia superpixels, as well the spatial relations among them via node edges. The experiments were conducted on real-world single polarization SAR images obtained from the Sentinel-1 satellite to test the performance of the proposed segmentation method. The results of these experiments show the advantage of the proposed GCN-based method for SAR image segmentation.

Deformation map prediction is a critical tool to foresee signs of abnormal events. Such forecasting facilitates quick countermeasure to avoid undesirable conditions. This work presents a novel recurrent neural network to forecast time-series deformation maps from InSAR data. Our proposed recurrent network employs a multi-scale attention mechanism to identify vital temporal features that influences subsequent deformation maps. We have evaluated our model on volcanic monitoring data using the Micronesia islands (Canary and Cape Verde archipelagos) Sentinel-1 imagery acquired between 2015 and 2018. The proposed method achieves minimal prediction error compared to the observed deformation values, suggesting the high reliability of our approach. The experimental results indicate the superiority of the proposed method in forecasting deformation maps with high accuracy compared to existing state-of-the-art approaches. Various ablation studies were conducted to study and validate the effectiveness of the multi-scale attention mechanism for deformation map forecasting.

The main objective of this study is to monitor the land infrastructure growth over a period of time using multimodality of remote sensing satellite images. In this project unsupervised change detection analysis using ITPCA (Iterated Principal Component Analysis) is presented to indicate the continuous change occurring over a long period of time. The change monitoring is pixel based and multitemporal. Co-registration is an important criteria in pixel based multitemporal image analysis. The minimization of co-registration error is addressed considering 8- neighborhood pixels. Comparison of results of ITPCA analysis with LRT (likelihood ratio test) and GLRT (generalized likelihood ratio test) methods used for SAR and MS (Multispectral) images respectively in earlier publications are also presented in this paper. The datasets of Sentinel-2 around 0-3 days of the acquisition of Sentinel-1 are used for multimodal image fusion. SAR and MS both have inherent advantages and disadvantages. SAR images have the advantage of being insensitive to atmospheric and light conditions, but it suffers the presence of speckle phenomenon. In case of multispectral, challenge is to get quite a large number of datasets without cloud coverage in region of interest for multivariate distribution modelling.

This work presents a system for multi-year crop type mapping based on the multi-temporal Long Short-Term Memory (LSTM) Deep Learning (DL) model and Sentinel 2 image Time Series (TS). The method assumes the availability of a pre-trained LSTM model for a given year and aims to update the corresponding crop type map fora different year considering a small amount of recent reference data. To this end, the proposed approach combines Self-Paced Learning (SPL) and fine-tuning (FT) techniques. While the SPL technique gradually incorporates samples from crop types that can be classified with high-confidence by the pre-trained model, the FT strategy adapts the network to those classes having low-confidence accuracy. This condition allows us to reduce the labeled samples required to achieve accurate classification results. The experimental results obtained on three tiles of the Austrian country on TSs of Sentinel 2 data acquired in 2019 and 2020 (considering a model pre-trained on images of 2018) demonstrate the capability of the LSTM to adapt to TS of images with different temporal and radiometric characteristic with respect to the one used to pre-train the model, with a relatively small number of training samples. As expected, by directly applying the model without performing any adaptation, we obtain a mean F-score (F1%) of 64% and 62% compared to 76% and 70% achieved by the proposed technique with only 1500 samples for 2019 and 2020, respectively.

This paper presents the proposal of a new change detection method for intensity VHF wavelength-resolution images. High-amplitude pixels are related to the presence of strong scatterers, resulting in high detection probability performance. However, the number of false alarms tends to be high too. In this initial study, difference images are considered to reduce the influence of the strong scatterers that are not related to targets, i.e., present in both surveillance and reference images. The proposed change detection method is based on a likelihood-ratio test, where the tested hypothesis is the bivariate exponential distribution. The derivation of the proposed likelihood test is presented. Finally, the proposed change detection method is assessed considering data measured with the CARABAS II VHF UWB SAR system. Preliminary results show that the proposed method is efficient in detecting positive changes.

In this paper, a problem of resource expenses needed for storage, processing and transferring a large number of high resolution digital remote sensing images is considered. Application of discrete atomic compression (DAC), which is an algorithm based on atomic wavelets, to solving this problem is studied. Dependence of efficiency of the DAC algorithm on its parameters, in particular, quality loss settings, a structure of discrete atomic transform, which is a core of DAC, and a method of quantized wavelet coefficients’ encoding, is investigated. Binary arithmetic coding and a combination of Huffman codes with run-length encoding are used to provide lossless compression of quantized atomic wavelet coefficients. Comparison of these methods is presented. A set of digital images of the European Space Agency is employed as test data. In addition, we discuss promising ways to improve the DAC algorithm.

In this paper, we consider a problem of lossy compression of three-channel or color images with application to remote sensing. The main task of such a compression is to provide a trade-off between compression ratio and quality of compressed data that should be appropriate for solving the basic tasks as classification of sensed terrains, object detection and so on. Then, alongside with a desire to increase compression ratio, one needs to control introduced distortions (image quality) to ensure that compressed data are appropriate for further use. We propose a way to lossy compression that is based on providing quality of compressed images not worse than desired according to quality metrics. The outcome of our approach is that classification accuracy either does not get worse than for uncompressed data (and sometimes even improves) or gets worse only slightly. Earlier, we have shown that it is possible to control quality for component-wise compression of multichannel images. Here, it is demonstrated that it is possible to control quality for 3D compression. Compared to component-wise compression, 3D approach leads to two important benefits: 1) compression ratio can be almost twice larger; 2) probability of correct classification can be slightly better. These benefits are confirmed for real-life three-component data acquired by Sentinel sensor using maximum likelihood-based classifier.

In this study, a technique for multifractal classification (MC) of synthetic aperture radar (SAR) images of ice-covered sea areas is proposed. This technique is based on the use of SAR image local Hölder exponents (LHEs) and coarse Hölder exponents (CHEs) calculated for sliding windows with different sizes. Hölder exponents are very effective and easily computable image texture descriptors that characterize the degree of local irregularity of image functions. The main steps of the presented SAR image classification technique are following: Sentinel-1 SAR image transformation (application of orbit file, radiometric calibration, speckle filtering, terrain correction and conversion to dB), extracting local and coarse Hölder exponents from HH- and HV-polarized Sentinel-1 SAR images, stacking local and coarse Hölder exponents into high-dimensional feature vectors, classification of the formed feature vectors by some classifier. Experimental testing of the proposed technique for classification of SAR images was carried out on several regions of Sentinel-1b SAR images demonstrating ice-covered areas of the Kara Sea. The first step of the technique was implemented by SNAP toolbox, and the next three steps were implemented using own MATLAB application (https://github.com/UchaevD/GMAToolbox). SAR image classification results were compared with ice charts of U.S. National Ice Center (NIC), which contain weekly information on sea ice concentration and ice thickness. As a result of comparison with NIC ice charts, it has been established that Kara Sea areas with widely-spreading types of floating ice can be successfully separated by MC of Sentinel-1b SAR images, and overall and average classification accuracies are not less than 75%. The results of the study suggest that MC of SAR images of ice-covered sea areas can be used to automate the generation daily ice charts for various ice-covered sea areas in the Arctic and Antarctic.

Radar sounders (RSs) mounted on airborne platforms are active sensors widely employed to acquire subsurface data of the cryosphere for Earth observation. RS data, also called radargrams, provide information on the buried geology by identifying dielectric discontinuities in the subsurface. Recently, a strong effort can be observed in designing automatic techniques to identify the main targets of the cryosphere. However, most of the methods are based on target-specific handcrafted features. Newly convolutional neural networks (CNNs) automatically extract meaningful features from data. However, supervised training requires numerous labeled data that are hard to retrieve in the RS domain. In this work, we adopt a CNN pre-trained in domains other than RS for automatically segmenting cryosphere radargrams. To adapt to the radargram characteristics, we introduce convolutional layers at the beginning of the pre-trained network. We modify the top layers of the network to a U- fashion autoencoder to extract relevant features for the target task. The new layers are fine-tuned with few labeled radargrams to identify and segment five targets: free space, continental ice layering, floating ice, bedrock, and EFZ and thermal noise. The pre-trained weights are not updated during fine-tuning. We applied the proposed approach to radargrams from Antarctica acquired by MCoRDS3, obtaining high overall accuracy. These results demonstrate the effectiveness of the method in segmenting radargrams and discriminating continental and coastal ice structures.

Radar sounders (RS) play an important role in planetary investigation. However, the complex tasks of predicting performance and interpreting RS data and predicting performance require to perform data simulations. In the literature, there are different methods for RS data simulation, including: i) numerical methods involving the exact solution of Maxwell's equations based on 3D modelling; and ii) methods based on ray-tracing and facet modelling. Among numerical methods, the Finite-Difference Time-Domain (FDTD) technique allows one to precisely model complex nonlinear targets, both geologically and geophysically, at the cost of high computational load. On the contrary, coherent ray-tracing methods reduce the computational cost by triangulating the targets into facets with size comparable to the wavelength. In this work, we combine the advantages of FDTD and ray-tracing methods into a novel integrated simulation technique for modelling scattering phenomena at two scales of wave interaction, i.e. large (facet) scale and small (sub-facet) scale. The coherent facet method is used to simulate facet- scale scattering phenomena, while FDTD is used to evaluate the sub-facet-scale scattering due to roughness on top of the single facets. We investigate the method's effectiveness by generating a one-layer synthetic DEM as a fractional Brownian motion (fBm) process superimposed by different values of RMS slope of the small-scale roughness and by analysing how the received signal changes in terms of signal-to-noise ratio. The results show the effectiveness of the method in representing small-scale roughness realistically.

In this study, we present spatial feature profiles that can be used in addition to spectral profiles for semisupervised small-sample-size (SSS) classification of hyperspectral images (HSIs). These profiles have been obtained by combining extended multi-attributive profiles (EMAPs) and the recently proposed Chebyshev moment multifractal profiles (CMMPs). In order to demonstrate SSS classification capabilities of the introduced feature profiles, several experiments were performed on two test HSIs. In these experiments, we used a graph-based ensemble learning method for semisupervised HSI classification and a small number of labeled samples for training. The experiments performed on test HSIs demonstrate that the proposed feature profiles provide good classification performance in terms of the overall accuracy (OA), average accuracy (AA) and Kappa statistics. We also compared the classification results obtained using EMAPs and CMMPs with those obtained using EMAPs alone, CMMPs alone, and another spatial feature sets. It has been established that the classification based on CMMPs and EMAPs shows obvious improvements, especially when the number of labeled training samples is extremely small. In the final part of the study, the HSI classification results obtained using the proposed feature profiles were compared with classification maps obtained by deep learning methods adopted for small training samples. This comparison showed that the semisupervised classification by EMAPs and CMMPs is characterized by higher values of OA, AA and Kappa coefficients. Moreover, in contrast to deep learning methods, the classification procedure based on the calculation of the proposed feature profiles and graph-based ensemble learning is not time-consuming.

The normalized difference vegetation index (NDVI) has been commonly used for vegetation monitoring, such as water stress detection, crop yield assessment, and evapotranspiration estimation. However, the influence of spatial resolution on the individual tree level NDVI using the Unmanned Aerial Vehicles (UAVs) is poorly understood. Therefore, in this research, the effects of the spatial resolution of UAV imagery are investigated using high-resolution multispectral images. A temporal sequence of UAV multispectral imagery was collected over an experimental pomegranate field, capturing variations in the whole growing season of 2019, at the USDAARS (U.S. Department of Agriculture, Agricultural Research Service) San Joaquin Valley Agricultural Sciences Center in Parlier, California, USA. The NDVI distribution of individual trees was generated at the 60 m, 90 m, and 120 m spatial resolution. Experimental results indicated how the spatial resolution of UAV imagery could affect NDVI values of individual trees.

Industrial storage tanks are widely used in petroleum, chemical industry, metallurgy and other process industries. The use of storage tanks are important in ensuring industrial safety production and product storage. Recently, concerns are rising over leaks and spills which leads to potential environmental and public health risks, for example, air pollution. Thus, there is a need to detect and monitor the tanks, which has become a regulatory requirement in most countries. The monitoring of industrial storage tanks in a city is of great practical significance to the planning and construction of the city. With the recent development of remote sensing technology, computer vision and image processing, it is possible to automatically detect industrial storage tanks from optical remote sensing images. This paper mainly studies the automatic detection of urban industrial storage tanks using deep learning based algorithm, aiming to help people manage and monitor the urban environment and resources. In this paper, we collected optical satellite remote sensing images of the city as a unit and created a city-level dataset including three cities of Guangdong province in south China by utilising satellite image data obtained from Google Earth imagery. To explore the effect of the deep learning object detection algorithm for the industrial storage tanks detection at the city-level, a deep learning model built with SSD (Single-Shot Detector) was trained based on our collected dataset. To further improve the detection accuracy of industrial storage tanks, Hough transform was used to reduce the false alarm in the deep learning results. The experimental results show that the combination of deep learning target detection model and Hough transform is effective in detecting of industrial storage tank on the collected city-level dataset and can achieve promising detection results.

In radar systems, when solving the problem of detecting objects behind dielectrically transparent obstacles against a background of noise and interference, methods of compressing modulated UWB signals are often used. The most widespread in solving this problem are Barker codes, which are binary sequences of finite length N = 2,3,4,5,7,11,13. One of the main features of Barker codes is the equality of the amplitudes of all lateral maxima of the autocorrelation function at their minimum possible level not exceeding 1 / N. Code sequences with such properties have not been found for N< 13. Recently, the theory of quasi-orthogonal matrices has arisen and is being developed, which include the Mersenne and Raghavararo matrices. Mersenne matrices exist in all orders N = 4n-1, where n is a natural number. The indicated matrices, which are the kernel of the Hadamard matrices and generalize them, can have both cyclic and symmetric constructions. In this paper, we consider the modulation of UWB signals using the Mersenne and Raghavararo codes obtained from the rows of the corresponding quasi-orthogonal matrices. The compression characteristics of code-modulated signals in comparison with Barker codes are investigated by the method of simulation modeling. The results of evaluating the compression characteristics of the considered signals showed the advisability of using, for example, for N = 13 the Raghavararo code instead of the Barker codes. This provides greater noise immunity of UWB signals in the channels for detecting objects behind obstacles. Since, in addition to radar systems, the Barker sequence of length 11 is widely used in digital data transmission systems, the developed simulation model and the results obtained using Mersenne codes are of great theoretical and practical importance in studies of noise immunity in digital UWB data transmission channels in a complex electromagnetic environment.

Ship detection from remote sensing images has been a topic of interest that gradually gained attention over the years due to the wide variety of its applications in the field of maritime surveillance, such as oil discharge control, sea pollution monitoring, and harbour management. Even though there is an extensive amount of methods developed for ship detection, there are still several challenges that remain unsolved, especially in complex environments. These challenges include occlusions due to shadows, clouds, and fog. Nowadays, deep learning algorithms, especially Deep Convolutional Neural Networks (DCNNs), are considered as a powerful approach for automatically detecting ships in satellite imagery. In this paper, enhanced Faster R-CNN (FRCNN) model will be used to overcome the aforementioned unsolved challenges. The enhanced FRCNN, which combines high level features with low level features, will be trained and tested in the frequency domain using the publicly available satellite imagery dataset, Airbus Ship Detection, provided by Kaggle. The performance will be compared to the original FRCNN based on their Overall Accuracy (OA) and Mean Average Precision (mAP) metrics.

Anomaly detection uses spectral pixels to distinguish between one pixel or group of pixels in a hyperspectral image and itstheir background pixels. Most of the anomaly detection algorithms depend on the assumptions of the background distribution such as the RX algorithm which assumes the gaussian distribution of the background which is not valid for most cases of hyperspectral images. Moreover, most of the algorithms have problems with the false alarms which is noise and detected as anomalies. To overcome these drawbacks, we propose a simple and easy anomaly detection algorithm which depends mainly on the spectral unmixing. Instead of using the raw pixels as given data to detect anomalies, we apply the spectral unmixing algorithm first to estimate the abundance maps and use these maps as features for anomaly detection. Next, we use edge detection algorithm for all abundance maps to detect all boundaries and anomalies in the scene. This gives robustness to the detection algorithm as every anomaly is detected in two abundance maps. We used AVIRIS hyperspectral imaging data cubes to evaluate the proposed algorithm.

Hyperspectral images (HSIs) provides abundant spectral information through hundreds of bands with continuous spectral information that can be used in land cover fine change detection (CD). HSIs make it possible for hyperspectral CD performance with higher discrimination on changes but provides a challenge to the conventional CD techniques due to its high dimensionality and dense spectral representation. In this paper, we implemented intrinsic image decomposition (IID) model to decompose the hyperspectral temporal difference image into two parts: real change and pseudo change information. In particular, the spectral reflecting component is selected as a kind of pure spectral feature used to enhance the CD performance in multitemporal HSIs. Experimental results illustrate the effectiveness of IID features extraction in addressing a supervised CD task.