Wetlands are vital for maintaining ecosystems and supporting biodiversity, but they face increasing threats from climate change and human activities. Accurate mapping of wetlands is essential to detect detrimental changes and guide effective conservation efforts. However, many wetland mapping (WM) methods using convolutional neural networks rely on kernels with fixed sizes and shapes, limiting their ability to capture the multiscale features of wetlands. To enhance their ability, we propose Multiscale Deformable DenseNet (MDD) by integrating deformable convolutions into the DenseNet architecture and employing a dual feature extractor. The deformable convolutions adapt kernel shapes and sampling to capture spatial patterns across scales, while the dual feature extractor uses varied kernel sizes for diverse receptive fields. These innovations significantly improve the classification accuracy for the complex WM task, where classes are often highly similar. Experimental results demonstrate that MDD achieves the highest overall accuracy (OA) in three hyperspectral data sets, with an OA of 97.24%, 98.23%, and 94.59%, compared to the best competing OAs of 96.33%, 97.16% and 92.70%, respectively. These results highlight MDD’s superiority in WM1.

Effective monitoring of wetlands plays a pivotal role in comprehending and managing these ecologically vital ecosystems. This study assesses the potential of C-band synthetic aperture radar (SAR) imagery in compact polarization (CP) mode, utilizing the RADARSAT Constellation Mission (RCM), for wetland characterization. We introduce the compact-polarimetric signature (CPS) as a novel descriptor to delineate wetlands, including bog, fen, and marsh classes. In addition, we propose an alternative decomposition technique ($\mu -\chi $) to segment the total power into three components: odd-bounce scattering$(P_{s})$, double-bounce scattering$(P_{d})$, and random scattering$(P_{v})$. For our evaluation, we selected a test site in New Brunswick, Canada, and acquired a series of RCM datasets covering this region. The time-series CPS plots yield valuable insights, elucidating the scattering mechanisms of different wetland classes. Notably, these plots reveal that during the active season, characterized by changing vegetation structures, the scattered waves exhibit variations, leading to changes in received power and the purity parameter ($\mu $). Furthermore, the observed variations in the proposed power components demonstrate a significant discriminatory capacity among wetlands. The$P_{s}$,$P_{d}$, and$P_{v}$components effectively distinguish bog, fen, and marsh classes, respectively, capturing the unique characteristics of each wetland type. These findings carry considerable potential for advancing wetland characterization through the RCM CP-SAR mission. The improved discriminative ability among different wetland classes is a valuable contribution to the broader field of wetland ecology and management. This advancement potentially empowers precise wetland classification, facilitating well-informed decision-making in wetland preservation and resource allocation. The applications of these findings extend to ecosystem monitoring, environmental impact assessments, and the long-term evaluation of wetland health. Eventually, this contributes to developing more effective wetland conservation and management strategies.

The main goal of this study is to investigate the potential of the Sentinel-1 sensor in accurately mapping different types of freshwater ice. The research was conducted at Saint-Pierre Lake in Quebec, Canada. We used a dataset that included backscatter data and grey level co-occurrence matrix (GLCM) properties. To classify the freshwater ice types, we employed the Fuzzy C-Means (FCM) unsupervised classification scheme. Furthermore, a feature selection analysis was done to identify the most relevant GLCM parameters that could effectively map the different ice types with accuracy. The preliminary findings demonstrate that incorporating both the Sentinel-1 VV polarization band and the GLCM Mean parameter leads to a significant improvement in distinguishing between the different types of ice cover.

This study contributes to advancing the understanding and utilization of compact polarimetry (CP) RCM Synthetic Aperture Radar (SAR) data for enhanced crop characterization and mapping. The received wave polarization signature captures the explicit variation of the received power with a fixed transmit polarization and varying received polarization bases. This information is then suitably utilized for improved discrimination among multiple crop types. Furthermore, the study explores using multi-date polarimetric features extracted from RCM imagery to achieve more accurate and detailed crop mapping results. By incorporating information from multiple acquisition dates, the multi-date polarimetric features illustrate excellent potential in capturing temporal variations in crop characteristics, leading to enhanced crop mapping accuracy. The implications and findings from this study could be essential in demonstrating the role of RCM data in agricultural applications.

In recent years, kernel-based methods and Deep Learning (DL) models have become the two most successful Remote Sensing (RS) analysis techniques for various Earth observations, particularly hyperspectral images. However, kernel-based methods are generally considered shallow models and intrinsically inconsistent with end-to-end learning. On the other hand, end-to-end learning is one of DL models' essential features as it seems to be responsible for their proven higher performances. Nevertheless, kernel methods are based on rigid mathematical theory and can efficiently cope with high-dimensional data. This paper proposed a hybrid deep kernel model to benefit from both kernel-based methods and DL models. This novel deep kernel model, namely Convolutional Kernel Network (CKN), was applied to two benchmark hyperspectral image datasets. Moreover, the proposed hybrid method was compared to Support Vector Machine (SVM) classifiers with various kernel functions. The experimental results indicated that the CKN's outperforms SVM.

The Canadian RADARSAT Constellation Mission (RCM) has passed its early operation phase, with the current performance evaluation. In this study, we provide results update on RCM performance for selected SAR applications. The RCM was designed to address three core applications-disaster management, maritime surveillance, and ecosystem monitoring. Our study shows a promising level of agreement between RCM and RADARSAT-2 performance in flood mapping using dual-polarized HH-HV SAR data over Red River, Manitoba. Visual analysis of coincident RCM compact polarimetric and RADARSAT-2 dual-polarized HH-HV SAR imagery over the Resolute Passage, Canadian Central Arctic, highlighted an improved contrast between sea ice classes in dry ice winter conditions. Object-oriented classification of a wetland area in Newfoundland and Labrador by fusion of RCM dual-polarized VV-VH data and Sentinel-2 optical imagery revealed promising classification results, with an overall accuracy of 91.1% and a kappa coefficient of 0.87.

We propose a hybrid algorithm for despeckling the Synthetic Aperture Radar (SAR) images using the Convolutional Neural Network (CNN) denoising and complex wavelet shrinkage. In particular, we perform the speckle reduction process in the complex wavelet domain. We first despeckled the approximation complex wavelet coefficients using the MUltichannel LOgarithm with the Gaussian denoising algorithm (MuLoG) based on a pre-trained CNN model named FFDNet. Next, we despeckled the log-transformed details of the complex wavelet coefficients using the averaged version of the Maximum a Posteriori (AMAP) estimator. The experimental results on simulated and real SAR images showed that the proposed method achieved better speckle suppression in the homogeneous areas while preserving edges and point targets than other state-of-the-art methods.

Shallow convolutional neural networks (CNNs) have successfully been used to classify polarimetric synthetic aperture radar (PolSAR) imagery. However, one drawback of the existing deep CNN-based techniques is that the input PolSAR training data are often insufficient due to their need for a significant number of training data compared to shallow CNN models utilized in PolSAR image classification. In this paper, we propose using Haar wavelet transform in deep CNNs for effective feature extraction to improve the classification accuracy of PolSAR imagery. Based on the results, the proposed deep CNN model obtained better average accuracy in the San Francisco region with an accuracy of 93.3% and produced more homogeneous classification maps with less noise compared to the two much shallower CNN models of AlexNet (87.8%) and a 2D CNN network (91%). The proposed algorithm is efficient and may be applied over large areas to support regional wetland mapping and monitoring activities using PolSAR imagery. The codes are available at (https://github.com/aj1365/DeepCNN_Polsar).

Leaf Area Index (LAI) and biomass are the most critical biophysical parameters for crop monitoring. In this study, we used three ensemble-based methods, including Random Forest (RF), Gradient Boosting (GB), and Extreme Gradient Boosting (XGB), for crop parameter estimation and mapping of soybean and wheat in an agricultural region in Winnipeg, Canada. Various Vegetation Indices (VIs) and radar parameters were extracted from multitemporal multispectral Sentinel-2 images and Synthetic Aperture Radar (SAR) Sentinel-1 data. Feature selection was made, first, based on the correlation between extracted features and target biophysical parameters. Features with low importance were then removed based on the correlation between all features. The RF model has the lowest RMSE among the examined methods for dry biomass, wet biomass, and LAI for soybean. For wheat, XGB has the lowest RMSE for dry and wet biomasses, while RF led to LAI's highest accuracy.

Wetlands are highly sensitive ecosystems that have experienced largely undocumented loss across Canada. Accurate statistics of historic loss of wetlands across many provinces is vague at best or non-existent at worst, as exemplified in Newfoundland and Labrador (NL). Thus, NL represents a perfect candidate for implementing historical remote sensing data sets and change detection methods. Given recent advancements in earth observation technology, it is now feasible to implement remote sensing-based change detection methods at scales never previously possible. As such, the goal of this work is to develop a methodology to assess wetland class change across the island of Newfoundland between 1985 and 2015 using historic and current Landsat imagery, Random Forest classification, and the Google Earth Engine (GEE) platform. The resulting accuracies ranged from 84.37% to 88.96%. The analysis reveals that wetland classes over the last 30 years have been unstable, and the biggest loss of wetlands to anthropogenic land cover occurred between the 1980's and the 1990's. Index Terms - Wetlands, Change Detection, Landsat, Geo big data

In this letter, we proposed a novel optimal view generation framework based on the genetic algorithm (GA) and Gaussian mixture models (GMMs) to improve multiview active learning (MV-AL). AL methods enlarge training data sets, by iteratively selecting the most informative samples, in order to improve the classification performance. By using multiple views to build multiple classifiers, the information content of each unlabeled samples can be more accurately estimated. The MV-AL methods are more inherently suitable for high-dimensional data such as hyperspectral images. This hybrid framework simultaneously constructs the optimal number of diverse and sufficient views. The proposed algorithm has two main steps. In the first step, by applying a cluster distortion function-based GMMs, the actual number of available independent views is determined. In the next step, a hybrid GA approach selects the optimal combination of views using two different criteria. The experiments were conducted on two benchmark hyperspectral data sets, namely, Kennedy Space Center (KSC) and Indian Pines AVIRIS. The results demonstrated an increase in diversity and sufficiency of the views compared to the traditional view generation methods. Furthermore, the performance of MV-AL has also been significantly improved.

This paper presents multiple kernel learning (MKL) in the context of remote sensing (RS) image classification problems by illustrating main characteristics of different MKL algorithms and analyzing their properties in RS domain. A categorization of different MKL algorithms is initially introduced, and some promising MKL algorithms for each category are presented. In particular, MKL algorithms presented only in machine learning are introduced in RS. Then, the investigated MKL algorithms are theoretically compared in terms of their: 1) computational complexities; 2) accuracy with different qualities of kernels; and 3) accuracy with different numbers of kernels. After the theoretical comparison, experimental analyses are carried out to compare different MKL algorithms in terms of: 1) model selection and 2) feature fusion problems. On the basis of the theoretical and experimental analyses of MKL algorithms, some guidelines for a proper selection of the MKL algorithms are derived.

Adoption of optical vegetation indices for local, national, and global crop condition monitoring is wide spread. Given that cloud cover impedes acquisition of these data, this research examines whether Synthetic Aperture Radar (SAR), specifically a compact polarimetric (CP) configuration, could augment these operational initiatives. Encouraging statistical correlations are reported between several CP parameters and the Normalized Difference Vegetation Index (NDVI). These early results suggest that further development is warranted to integrate a SAR-based index with optical-NDVI particularly considering the configuration of future Canadian satellite systems.

Rational polynomial coefficients (RPCs) biases and over-fitting phenomenon are two major issues in terrain-independent rational function models. These problems degrade the accuracy of extracted spatial information from very high spatial resolution (VHSR) satellite images. This study particularly focused on overcoming the over-fitting problem through an optimal term selection approach. To this end, multiobjective genetic algorithm was used in order to optimize three effective objective functions: the RMSE of ground control points (GCPs), the number, and the distribution of both RPCs and GCPs. Finally, the technique for order of preference by similarity to ideal solution, as an efficient multicriteria decision-making method, was applied to select the best solution, i.e., the optimum terms of RPCs, through the ranking of solutions in the optimum set. The performance of the proposed method was evaluated by using three VHSR images acquired by GeoEye-1, Worldview-3, and Pleiades satellite sensors. Experimental results show that subpixel accuracy can be nearly achieved in all data sets, when over-fitting problem is addressed. The optimal selected terms leaded to a significant improvement compared to the original RPCs. Indeed, our method, which is independent of GCPs distribution, not only requires a small number of GCPs, but also leads to a 30% to 75% improvement when compared to the original RPCs. This improvement in VHSR images, usually makes no more need to remove the RPCs biases.

This paper analyzes and compares different Multiple Kernel Learning (MKL) algorithms for the classification of remote sensing (RS) images. The main purpose of the comparison is to identify advantages and disadvantages of different MKL algorithms in terms of their computational time and classification accuracy. Furthermore, some guidelines on the proper selection of the MKL algorithms associated with different RS image classification problems are derived.

The National Aeronautics and Space Administration (NASA) Soil Moisture Active Passive (SMAP) satellite is scheduled for launch in January 2015. In order to develop robust soil moisture retrieval algorithms that fully exploit the unique capabilities of SMAP, algorithm developers had identified a need for long-duration combined active and passive L-band microwave observations. In response to this need, a joint Canada-U.S. field experiment (SMAPVEX12) was conducted in Manitoba (Canada) over a six-week period in 2012. Several times per week, NASA flew two aircraft carrying instruments that could simulate the observations the SMAP satellite would provide. Ground crews collected soil moisture data, crop measurements, and biomass samples in support of this campaign. The objective of SMAPVEX12 was to support the development, enhancement, and testing of SMAP soil moisture retrieval algorithms. This paper details the airborne and field data collection as well as data calibration and analysis. Early results from the SMAP active radar retrieval methods are presented and demonstrate that relative and absolute soil moisture can be delivered by this approach. Passive active L-band sensor (PALS) antenna temperatures and reflectivity, as well as backscatter, closely follow dry down and wetting events observed during SMAPVEX12. The SMAPVEX12 experiment was highly successful in achieving its objectives and provides a unique and valuable data set that will advance algorithm development.

Many researches have demonstrated that the spatial information can play an important role in the classification of hyperspectral imagery. Recently, an effective approach for spectral-spatial classification has been proposed using Minimum Spanning Forest (MSF) grown from automatically selected markers. This paper aims at improving this approach for classification of hyperspectral images in urban areas. The proposed framework is based on deriving the optimal spatial and spectral features of the original hyperspectral image using the Principal Component Analysis (PCA). The spatial features extracted in this study are wavelet, Gabor filter, mean, contrast, entropy, variance, homogeneity, dissimilarity, second moment, and correlation. The experimental results on three hyperspectral datasets demonstrate that compared to the original MSF-based approach, the proposed framework yields more accurate classification maps.

This paper presents a hyperspectral image classification method based on the semi-supervised random forest (SSRF) algorithm. This method uses Deterministic Annealing (DA) and the random forest classifier (RFC). The first step consists of performing the random forest algorithm by using labeled data. Then, image is classified and the probability of each unlabeled data will be computed. Based on the probability and the temperature parameter, label of unlabeled data will be determined. Finally, the classification is carried out based on the labeled data and unlabeled data which were converted to labeled data in the procedure of algorithm. The proposed method and also a conventional RFC method have been applied to an APEX (Airborne Prism Experiment) hyperspectral image. The results show more consistency in homogeneous area. In addition, its overall accuracy of classification is 82.63%, while the kappa coefficient is 0.78, and both are higher than the accuracies of spectral based classification using the conventional RFC, i.e. 73.58% and 0.68 respectively.

This paper presents a multi-step framework for classification and crop mapping using several polarimetric features, extracted from multitemopral Synthetic Aperture Radar (SAR) imagery. The multi-temporal data classification, not only improves the overall retrieval accuracy, but also provides more reliable crop discrimination in comparison to single-date data [1]. This is mainly because various phenogical stages of crops can contribute discrimination and classification of agricultural lands. The proposed framework in this paper consists of three main steps: a) data preprocessing, b) processing, and c) classification and evaluation. Several polarimetic features are extracted from preprocessed data, including the coherency and/or the covariance matrixes. Polarimetry decompositions then can allpy to ectract the statistical or physical based polarimetric components. Support vector machines' (SVM) classifier is employed for classification of these features. In addition, different kinds of kernel functions are used to evaluate the performance of SVM for classification. The method is applied to several UAVSAR L-band SAR images acquired over an agricultural area near Wennipeg, Manitoba, Canada. in summer of 2012. The experimental tests show that using two data data increases the overall accuracy of the classification up to 14%, and using an aditional date, i.e. three multitemporal datasets, increases the overall accuracy about 9% in comparing to two date imagery. The effect of multi-temporal data in crop classification is much more than even using more training data, which sometimes is expensive and time consuming.

The accurate estimation of the number of endmembers (NOE) in a given hyperspectral imagery plays a fundamental role in the effective classification, clustering, unmixing, and identification of the materials presenting in any remote scene. The optimal estimation of the NOE, however, is a quite challenging task, due to the inevitable combined presence of noise and outliers. In the last decade, several algorithms have been proposed to estimate the exact NOE. Nonetheless, these methods usually lead to different values for intrinsic dimensionality. These uncertainties make the user unable to determine the right intrinsic dimension. This letter proposes a statistical based method for finding the NOE in hyperspectral imagery. In the first step of this method, a number of candidates are selected using the residual analysis and change-point analysis. Then, according to application, one of these candidates can be selected. For this selection, here, an intrinsic dimensionality estimator, based on the singular value decomposition (SVD), is used to make this selection. Based on a comparison with second moment linear and outlier-geometry based estimation of NOE-affine hull (O-GENE-AH), the proposed method yields better results.

Land covers and uses are dynamically being changed over the time. Detection and identification of these changes is necessary and is the first step of any study or planning for natural resource management. Synthetic Aperture Radar (SAR) imagery, thanks to its independence to weather conditions and sun illumination, is a powerful tool for these studies. In this research an unsupervised change detection framework based on the kernel-based clustering technique is presented. Kernel C-means algorithm is employed to separate the changes classes from the no-changes. This method is a non-linear algorithm which considers the contextual information. Using the kernel functions, the projecting of the data into a higher dimensional space helps to make the non-linear features more separable in a linear space. The proposed methodology has applied to dual-pol L-band SAR images acquired by the ALOS from Urmia Lake. Results show because of non-linear behavior of changed phenomenon, the algorithm leads to more reliable results.

Synthetic Aperture Radar (SAR) satellite sensors recently provide valuable sources of earth observation data for various environmental applications. Beside the specifics properties of these data including multi-polarization and polarimetric image data, the presence of unavoidable speckle seriously degrades the quality of these data. Specifically, in certain applications such as clustering, classification and change detection speckles make some difficulties in analysis data and interpretation of results. In this research, a hybrid approach, based on frequency-domain transforms, is proposed. This method is a combination of wavelet and curvelet transforms to suppress the speckle noise in SAR images. This approach based on features and region which has a good efficiency in removing noise and preserving information of data in case of edges and shape. Results of these methods were compared simultaneously and with conventional speckle filtering methods (e.g. Lee, Frost and Kuan).

In recent studies, the support vector data description (SVDD) has been successfully applied to the classification of hyperspectral images. However, there is a major problem with this approach, namely, the precise setting of the Gaussian kernel width (i.e., the sigma), which is, in fact, the common limitation of kernel methods in achieving a reliable performance. Generally, the sigma is tuned for multiclass data sets through theK-fold cross validation (KCV), a time-consuming method. To reduce the computation time in real-time applications, typically, theKCV is used to constrain all the involved SVDD classifiers to share the same sigma. This letter presents a fast and straightforward method to estimate the sigma for each individual SVDD classifier based on statistical properties of the Gaussian kernel. To evaluate the performance of the proposed method, three frequently used hyperspectral data sets are employed. The results are then compared to theKCV method for sigma selection, and, in addition, two direct sigma estimation methods. Preliminary results using incomplete training data suggest that the proposed method can achieve similar or better performance with faster processing times than theKCV and also provide a significant superior performance in comparison with the direct methods.

The general objective of anomaly detection (AD) in hyperspectral imagery is to detect full-pixel targets. To meet this purpose, the global AD methods can achieve more reliable results than the local methods in terms of time and accuracy. The kernel-based Support Vector Data Description (SVDD) has recently received great attention in the hyperspectral AD applications. This paper presents a global SVDD-based method for autonomous full-pixel AD. The method consists of three steps: clustering, background modeling, and autonomous AD. Experimental results on a hyperspectral dataset show the superiority of the proposed method comparing to the global based SVDD method.

Recently, anomaly detection (AD) has attracted considerable interest in a wide variety of hyperspectral remote sensing applications. The goal of this unsupervised technique of target detection is to identify the pixels with significantly different spectral signatures from the neighboring background. Kernel methods, such as kernel-based support vector data description (SVDD) (K-SVDD), have been presented as the successful approach to AD problems. The most commonly used kernel is the Gaussian kernel function. The main problem using the Gaussian kernel-based AD methods is the optimal setting of sigma. In an attempt to address this problem, this paper proposes a direct and adaptive measure for Gaussian K-SVDD (GK-SVDD). The proposed measure is based on a geometric interpretation of the GK-SVDD. Experimental results are presented on real and synthetically implanted targets of the target detection blind-test data sets. Compared to previous measures, the results demonstrate better performance, particularly for subpixel anomalies.