Deep learning models, widely used for their high accuracy in various applications such as remote sensing image classification, are frequently seen as black boxes due to their complex internal workings. Explainable artificial intelligence, a recent field of research, seeks to clarify the decision-making processes of these deep learning models, making them more transparent and comprehensible. In this study, a new model-agnostic explainable method based on sensitivity analysis is proposed. This method works by observing how the model’s prediction changes when different parts of an image are perturbed using the meta-model representation. High Dimensional Model Representation is utilized as a meta-model due to its strengths in model approximation capability with a few feature interactions, allowing for efficient analysis and understanding of complex models with reduced computational complexity. The proposed approach is applied to a convolutional neural network model, specifically for tackling the remote sensing scene classification challenge using the EuroSAT dataset. The results are compared to those of the LIME method.

Cotton, a vital crop in the global textile industry, faces challenges from climate and ecosystem changes. Accurate cotton yield prediction is crucial for the economy and environmental sustainability, and it requires a deep understanding of the complex relationship between its parameters and yield. To achieve this, a comprehensive approach integrating climatic factors, soil parameters, and biophysical parameters observed through high-resolution remote sensing satellites was employed. This study utilized a multisource dataset to develop a predictive model for cotton yield over Turkiye, allowing accurate yield estimation and understanding the impact of the Earth Observation (EO)-based yield predictors on the model. Specifically, we utilized the Explainable Boosting Machine (EBM) algorithm to model and predict cotton yield while offering insights into selecting EO predictors. Additionally, we conducted a performance evaluation of our proposed approach in comparison to popular boosting-based algorithms like eXtreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), and Light gradient boosting (Light-GBM).

This study addresses the explainability challenges of deep-learning models in the context of hyperspectral remote sensing image classification. Three prominent explainable artificial intelligence methods, namely GradCAM, GradCAM++, and Guided Backpropagation, have been employed in order to comprehend the decision-making process of a typical convolutional neural network model during spatial-spectral hyperspectral image classification. The experiments that have been conducted investigate the impact of pixel patch sizes on spatial attention, as well as spectral band importance. The findings provide insights into the behavior of both convolutional neural networks, as well as the comparative performance of explainability techniques.

Cotton is under the threat of climate and ecosystem change, and has an essential role in the global textile industry. This makes its yield prediction essential for both economics and sustainability. The potential cotton yield can be predicted by integrating climatic factors, soil parameters, and biophysical parameters observed by high temporal & spatial resolution remote sensing satellites. This study used a multisource dataset to create an explainable and accurate predictive model for cotton yield prediction over the continental US (CONUS). A recently proposed glass-box method called Explainable Boosting Machine (EBM), which provides transparency, reliability, and ease of interpretation, was implemented. Accuracy performance was compared with common machine learning (ML) methods for predicting cotton yields. The EBM showed higher accuracy against other glass-box methods and competitive results with black-box models. With the help of the EBM, the importance of individual features and their pairwise interactions was revealed without applying any post-hoc methods. The study findings showed that the precipitation (P), enhanced vegetation index (EVI), and leaf area index (LAI) are the three most important dynamic features. The dynamic features are the driver of the created model with 78% of the overall feature importance, followed by pairwise interactions of the features with 16% contribution. Lastly, static features contribute 6% to the overall feature importance. The study highlights the importance of using multi-source data and interactions of the input features and providing an interpretable model to understand the inner dynamics of cotton yield predictions.

Feature selection is essential in various fields of science and engineering, from remote sensing to computer vision. Reducing data dimensionality by removing redundant features and selecting the most informative ones improves machine learning algorithms’ performance, especially in supervised classification tasks, while lowering storage needs. Graph-embedding techniques have recently been found efficient for feature selection since they preserve the geometric structure of the original feature space while embedding data into a low-dimensional subspace. However, the main drawback is the high computational cost of solving an eigenvalue decomposition problem, especially for large-scale problems. This paper addresses this issue by combining the graph embedding framework and representation theory for a novel feature selection method. Inspired by the high dimensional model representation, the feature transformation is assumed to be a linear combination of a set of univariate orthogonal functions carried out in the graph embedding framework. As a result, an explicit embedding function is created, which can be utilised to embed out-of-samples into low-dimensional space and provide a feature relevance score. The significant contribution of the proposed method is to divide ann-dimensional generalised eigenvalue problem intonsmall-sized eigenvalue problems. With this property, the computational complexity of the graph embedding is significantly reduced, resulting in a scalable feature selection method, which could be easily parallelized too. The performance of the proposed method is compared favourably to its counterparts in high-dimensional hyperspectral image processing in terms of classification accuracy, feature stability, and computational time.

Recent studies in remote sensing reveal that complex nonlinear learning models such as deep learning or ensemble-based learning are mostly preferred compared to shallow machine learning methods in solving many problems such as classification, image fusion, change detection, unmixing, and object recognition. The fact that much remote sensing data can be obtained quickly, abundantly, and free of charge, and the increasing computing power of computers with developing technology, are why such methods are preferred. With the emergence of big data, these methods provide more effective solutions than in past years, and they can outperform shallow machine learning methods in many remote sensing applications. Despite their high accuracy, such learning models have several limitations due to their black-box structure. Because of the high nonlinearity in predictive models, these models cannot explain why and how decisions are made. This paper presents a global model distillation approach to replace a black-box model with a fully explainable surrogate model utilizing polynomial chaos expansion. Preliminary results show that the proposed method can accurately replace a complex nonlinear model with a simpler one in hyperspectral image classification.

Feature selection has been a prominent research topic for a long time, not only in hyperspectral image classification but also in other related fields. It has gained even more popularity recently, especially with the growing interest in explainable AI studies. The literature on feature selection is extensively studied not only in remote sensing but also in the domain of computer science. However, most of the conventional approaches ignore information about the manifold structure of the data, which might be critical, especially for the analysis of hyperspectral data due to their complex nonlinear structure. This study introduces a feature selection approach based on graph embedding and global sensitivity analysis, utilizing the first-order terms of the high dimensional model representation. The effectiveness of the proposed method is analyzed on four hyperspectral datasets utilizing some evaluation criteria, including classification accuracy and clustering quality, and compared to seven state-of-the-art feature selection methods. The results show that the proposed method typically outperforms the others and is notably more computationally efficient.

Learning the manifold structure of remote sensing images is of paramount relevance for modeling and understanding processes, as well as encapsulating the high dimensionality in a reduced set of informative features for subsequent classification, regression, or unmixing. Manifold learning methods have shown excellent performance when dealing with hyperspectral image (HSI) analysis, but, unless specifically designed, they cannot provide an explicit embedding map readily applicable to out-of-sample (OOS) data. A common assumption to deal with the problem is that the transformation between the high-dimensional input space and the latent space (typically low) is linear. This is a particularly strong assumption, especially when dealing with HSIs due to the well-known nonlinear nature of the data. To address this problem, a manifold learning method based on high-dimensional model representation (HDMR) is proposed, which enables a nonlinear embedding function to project OOS samples into the latent space. The proposed method is compared to manifold learning methods along with their linear counterparts and achieves promising performance in terms of classification accuracy for a representative set of HSIs.

Hyperspectral images are the multidimensional matrices consisting of hundreds of spectral feature vectors. Thanks to these large number of features, the objects on the Earth having similar spectral characteristics can easily be distinguished from each other. However, the high correlation and the noise between these features cause a significant decrease in the classification performances, especially in the supervised classification tasks. In order to overcome these problems, which is known in the literature as Hughes's effects or curse of dimensionality, dimensionality reduction techniques have frequently been used. Feature selection and feature extraction methods are the ones used for this purpose. The feature selection methods aim to remove the features, including high correlation and noise, out of the original feature set. In other words, a subset of relevant features that have the ability to distinguish the objects is determined. The feature extraction methods project the high dimensional space into a lower-dimensional feature space based on some optimization criterion, and hence they distort the original characteristic of the dataset. Therefore, the feature selection methods are more preferred than the feature extraction methods since they preserve the originality of the dataset. Based on this motivation, an evolutionary based optimization algorithm utilizing self organizing map was accordingly modified to provide a new feature selection method for the classification of hyperspectral images. The proposed method was compared to well-known feature selection methods in the classification of two hyperspectral datasets: Botswana and Indian Pines. According to the preliminary results, the proposed method achieves higher performance over other feature selection methods with a very less number of features.

It is significant to monitor the phenological stages of agricultural crops with accurate and up-to-date information. In monitoring the phenological phases of some crops, optical remote sensing data offers significant spectral information and outstanding feature identification. However, a continuous time series of optical remote sensing data is difficult to obtain due to the weather dependency of optical acquisitions. In this paper, the feasibility of transfer learning between the features of Sentinel-1 and Sentinel-2 is evaluated to reduce these difficulties. A feature translation based on deep learning (DL) method, namely Cycle-Consistent Generative Adversarial Networks (cycle-GAN), was applied between Sentinel-1 and Sentinel-2 data. In order to evaluate the effect of the cycle-GAN method on crop type mapping and identification, Random Forest classification was applied to four different cases (Real SAR, Fake Optical + Real SAR, Real Optical, and Real Optical + Real SAR).

Manifold learning methods are very efficient methods for hyperspectral image (HSI) analysis but, unless specifically designed, they cannot provide an explicit embedding map readily applicable to out-of-sample data. A common assumption to deal with the problem is that the transformation between the high input dimensional space and the (typically low) latent space is linear. This is a particularly strong assumption, especially when dealing with hyperspectral images due to the well-known nonlinear nature of the data. To address this problem, a manifold learning method based on High Dimensional Model Representation (HDMR) is proposed, which enables to present a nonlinear embedding function to project out-of-sample samples into the latent space. The proposed method is compared to its linear counterparts and achieves promising performance in terms of classification accuracy of hyperspectral images.

For an accurate earthquake damage assessment from very high resolution (VHR) images, contextual relations between pixels need to be included in conjunction with spectral information during the classification. To utilize the spatial information in an efficient way, specific patterns representing the earthquake-induced damage should properly be modelled. Attribute Profiles (APs) and Multi Attribute Profiles (MAPs) provide a multi-dimensional representation of an image with a successive implementation of different attribute filters, and they are able to generate the complicated features for a specific pattern. In this study, the APs and the MAPs were used for the first time to extract the additional contextual features from very high resolution satellite image of City of Bam (Iran) acquired eight days after the earthquake. The performance of the morphological attribute features was compared to the those of Haralick's features (HFs) using the k-nn classifier, and the preliminary results showed that the APs and MAPs detect the earthquake damage more accurate than the HFs.

Dimensionality reduction has been an active research topic in hyperspectral image analysis due to complexity and non-linearity of the hundreds of the spectral bands. Locality preserving projection (LPP) is a linear extension of the manifold learning and has been very effective in dimensionality reduction compared to linear methods. However, its performance heavily depends on construction of the graph affinity matrix, which has two parameters need to be optimized: k-nearest neighbor parameter and heat kernel parameter. These two parameters might be optimally chosen simply based on a grid search when using only one representative kernel parameter for all the features, but this solution is not feasible when considering a generalized heat kernel in construction the affinity matrix. In this paper, we propose to use heuristic methods, including harmony search (HS) and particle swarm optimization (PSO), in exploring the effects of the heat kernel parameters on embedding quality in terms of classification accuracy. The preliminary results obtained with the experiments on the hyperspectral images showed that HS performs better than PSO, and the heat kernel with multiple parameters achieves better performance than the isotropic kernel with single parameter.

The role of global sensitivity analysis (GSA) is to quantify and rank the most influential features for biophysical variable estimation. In this letter, an approximation model, called high-dimensional model representation (HDMR), is utilized to develop a regression method in conjunction with a GSA in the context of determining key input drivers in the estimation of crop biophysical variables from polarimetric synthetic aperture radar data. A multitemporal Radarsat-2 data set is used for the retrieval of three biophysical variables of barley: leaf area index, normalized difference vegetation index, and Biologische Bundesanstalt, Bundessortenamt und CHemische Industrie stage. The HDMR technique is first adopted to estimate a regression model with all available polarimetric features for each biophysical parameter, and sensitivity indices of each feature are then derived to explain the original space with a smaller number of features in which a final regression model is established. To evaluate the applicability of this methodology, root-mean square and coefficient of determination were performed under different amounts of samples. Results highlight that HDMR can be used effectively in biophysical variable estimation for not only reducing computational cost but also for providing a robust regression.

Unsupervised manifold learning has become accepted as an important tool for reducing dimensionality of a dataset by finding its meaningful low-dimensional representation lying on an unknown nonlinear subspace. Most manifold learning methods only embed an existing dataset, but do not provide an explicit mapping function for novel out-of-sample data, thereby potentially resulting in an ineffective tool for classification purposes, particularly for iterative methods such as active learning. To address this issue, out-of-sample extension methods have been introduced to generalize an existing embedding of new samples. In this work, a novel out-of-sample method is introduced by utilizing High Dimensional Model Representation (HDMR) as a nonlinear multivariate regression with the Tikhonov regularizer for unsupervised manifold learning algorithms. The proposed method was extensively analyzed using illustrative datasets sampled from known manifolds. Several experiments with 3D synthetic datasets and face recognition datasets were also conducted, and the performance of the proposed method was compared to several well-known out-of-sample methods. The results obtained with Locally Linear Embedding (LLE), Laplacian Eigenmaps (LE), and t-Distributed Stochastic Neighbor Embedding (t-SNE) showed that the proposed method achieves competitive even better performance than the other out-of-sample methods.

Tracking crop's biophysical parameters using temporal Pol-SAR (Polarimetric Synthetic Aperture Radar Data) data is an active research topic in precision agriculture due to the sensitivity of PolSAR acquisition to canopy's physical and geometrical structure. Reconstruction of polarimetric features from collection of SAR data is computationally expensive, and more important, the inter-features correlations cause decreased performance in regression based biophysical parameter estimation. With the scope of operational crop monitoring, this study provides key variables to drive Leaf Area Index (LAI) from polarimetric data based on global sensitivity analysis (GSA) addressing the ranking of the most influential features. We applied variance-based GSA for temporal fully-polarimetric RadarSAt-2 images acquired through the cultivation period of two crops; canola and barley. Among 20 polarimetric features, anisotropy and correlation magnitude between co-polar channels were found to be the most influential polarimetric features for canola and barley, respectively.

Hyperspectral image classification via supervised approaches is often affected by the high dimensionality of the spectral signatures and the relative scarcity of training samples. Dimensionality reduction (DR) and active learning (AL) are two techniques that have been investigated independently to address these two problems. Considering the nonlinear property of the hyperspectral data and the necessity of applying AL adaptively, in this paper, we propose to integrate manifold and active learning into a unique framework to alleviate the aforementioned two issues simultaneously. In particular, supervised Isomap is adopted for DR for the training set, followed by an out-of-sample extension approach to project the large amount of unlabeled samples into previously learned embedding space. Finally, AL is performed in conjunction with k-nearest neighbor (kNN) classification in the embedded feature space. Experiments on a benchmark hyperspectral dataset illustrate the effectiveness of the proposed framework in terms of DR and the feature space refinement.

Unsupervised manifold learning has become accepted as an important tool for reducing dimensionality of a data set by finding its meaningful low dimensional representation lying on an unknown nonlinear subspace. Most manifold learning methods only embed an existing data set, but do not provide an explicit mapping function for novel out-of-sample data, thereby potentially resulting in an ineffective tool for classification purposes. To address this issue, out-of-sample extension methods have been introduced to generalize an existing embedding to new samples. In this work, a meta-modelling method called High Dimensional Model Representation (HDMR) is firstly implemented as a nonlinear multivariate regression for the out-of-sample problem for non-parametric unsupervised manifold learning algorithms. Several experiments show that the proposed method outperforms several state-of-the-art out-of-sample extension methods in terms of generalization to new samples for classification experiments on two remote sensing hyperspectral data sets.

In hyperspectral image analysis, the classification task has generally been addressed jointly with dimensionality reduction due to both the high correlation between the spectral features and the noise present in spectral bands, which might significantly degrade classification performance. In supervised classification, limited training instances in proportion with the number of spectral features have negative impacts on the classification accuracy, which is known as Hughes effects or curse of dimensionality in the literature. In this paper, we focus on dimensionality reduction problem, and propose a novel feature-selection algorithm, which is based on the method called high dimensional model representation. The proposed algorithm is tested on some toy examples and hyperspectral datasets in comparison with conventional feature-selection algorithms in terms of classification accuracy, stability of the selected features and computational time. The results show that the proposed approach provides both high classification accuracy and robust features with a satisfactory computational time.

Support vector machines is a very popular method in classification of hyperspectral images due to their good generalization capability even with a limited number of training datasets. However, the performance of SVM strongly depends on selection of kernel parameters when RBF kernel is used. In order to achieve a high classification performance, the kernel parameters, that are the value of regularization term and kernel width, should optimally be chosen. In this work, the use of recently developed evolutionary optimization methods, harmony search and differential evolution methods, are investigated in the context of hyperspectral image classification for the first time in this paper. The experimental results showed that these methods provide fast and accurate results in comparison to classical grid search approach.

Nowadays, hyperspectral images have been an attractive subject for many researches in remote sensing area since they provide abundant information due to their wide range of spectral bands. On the one hand, classification plays a significant role in extraction of information for different applications. On the other hand, providing a huge amount of data by hyperspectral images may lead to complexity and bring some redundancy due to high correlation among the hyperspectral bands. In order to reduce the redundancy, feature selection algorithms have been carried out to remove irrelevant features to efficiently use the classifier and to achieve a significant accuracy with minimum costs. In this work, a comprehensive analysis of well known feature selection algorithms will be conducted with different classifiers on some commonly used hyperspectral datasets. The contribution of this paper is to present an extensive benchmark study on using feature selection algorithms with hyperspectral dataset. The analysis of feature selection algorithms will be carried out by considering number of training samples, classification accuracy and computational time.

In hyperspectral image analysis, the classification task has generally been discussed with dimensionality reduction due to high correlation and noise between the spectral features, which might cause significantly low classification performance. In supervised classification, limited training samples in proportion to the number of spectral features have also negative impacts on the classification accuracy, which has known as Hughes effects or curse of dimensionality in the literature. In this paper, we focus on dimensionality reduction problem, and proposed a novel feature selection algorithm by using the method called random sampling high dimensional model representation (RS-HDMR), and the proposed algorithm were tested on a toy and hyperspectral dataset in comparison to conventional feature selection algorithms with regards to both computational time and classification accuracy.

The crop phenology retrieval on precision agriculture has been an important research area with the increasing demand on crops. Remotely sensed Synthetic Aperture Radar (SAR) data provides a simple possibility for automatic monitoring of agricultural fields due to the its inherit all-weather monitoring capability. Most of the studies rely on morphology based modelling of the electromagnetic backscattering which requires Monte Carlo simulations. In this paper, instead of modelling the backscattering of the signals for monitoring the crop fields, a classification scheme was implemented on the data acquired by TerraSAR-X by using the features extracted from backscattering coefficients with the machine learning algorithms which are Support Vector Machines, k-Nearest Neighbor and Regression Tree.

In classification of hyperspectral image, a common challenge is to deal with Hughes phenomenon also known curse of dimensionality, which is caused by high dimension with low samples and resulting in a poor classification performance [1]. There have been many ongoing researches in the literature to mitigate the Hughes phenomenon and accordingly increase the classification performance [2], [3], [4]. Support vector machines (SVM) is the one of the most important algorithm used in the classification of hyper-spectral image which is generally not effected by curse of dimensionality. Although it provides a good generalization ability in classification of hyperspectral dataset, recently, in order to increase the performance of SVM with the limited training data, a recursive feature elimination (RFE) approach based on SVM classifier has been introduced in order to rank the features with respect to their contribution to classification performance [5]. RFE approach utilize the objective function as a feature ranking criterion in order to eliminate the redundant features, and to produce a list of features having more discriminant ability. The experiments in the hyperspectral data classification by SVM also showed that the SVM-RFE method does not affected from the curse of dimensionality even if the number of samples are limited, and the satisfactory classification performance is obtained with using a small number of features [6].

In the recent years, new techniques so called non-parallel support vector machines (NSVM) have been developed and applied to some synthetic and UCI machine learning data sets, yielding competitive results especially in terms of computational complexity and classification performance compared to classical SVM [1]. In binary classification task, the aim of NSVM is to find two non parallel hyperplanes such that each plane is as close as possible to one of the two classes and also as far as possible from the other class. The study of NSVM algorithms was first began with proximal SVM classification which generates two parallel hyperplanes [2]. Afterwards, it has been demonstrated by several different approaches that classification problems could also be tackled with the use of non-parallel hyperplanes. The first non-parallel hyperplane classifier was introduced by Mangarisan and Wild (2006) named as the generalized eigenvalue proximal support vector machine (GEPSVM) [3]. They removed/dropped the parallelism condition of the generated planes and make the first plane be located as close as possible to one data set while keeping it furthest from the points of the other data set and vice versa. Each proximal plane was found by solving two generalized eigenvalue problems instead of solving a quadratic programming problem as it was required for classical Support Vector Machine. In some cases, better classification accuracy results were achieved with GEPSVM in a short span of time compared to classical support vector machine classification algorithms [4].

As the number of feature increases, classification accuracy may decrease. Additionally, computational overload increases with a large number of features. For effective classification performance and shortened the training time, the redundant features should be eliminated before the classification process. In this paper, a new HDMR-based feature selection approach is presented, sorting the features with respect to their sensitivity coefficient calculated by HDMR sensitivity analysis. With the experiments conducted, the HDMR-based feature selection approach is competitive with sequential forward feature selection method and faster in terms of computational time, especially when dealing with datasets having a large number of features.

Classification of nonlinearly separable data by nonlinear support vector machines (SVMs) is often a difficult task, particularly due to the necessity of choosing a convenient kernel type. Moreover, in order to get the optimum classification performance with the nonlinear SVM, a kernel and its parameters should be determined in advance. In this paper, we propose a new classification method called support vector selection and adaptation (SVSA) which is applicable to both linearly and nonlinearly separable data without choosing any kernel type. The method consists of two steps: selection and adaptation. In the selection step, first, the support vectors are obtained by a linear SVM. Then, these support vectors are classified by using the$K$-nearest neighbor method, and some of them are rejected if they are misclassified. In the adaptation step, the remaining support vectors are iteratively adapted with respect to the training data to generate the reference vectors. Afterward, classification of the test data is carried out by 1-nearest neighbor with the reference vectors. The SVSA method was applied to some synthetic data, multisource Colorado data, post-earthquake remote sensing data, and hyperspectral data. The experimental results showed that the SVSA is competitive with the traditional SVM with both linearly and nonlinearly separable data.

A linear support vector machine (LSVM) is based on determining an optimum hyperplane that separates the data into two classes with the maximum margin. The LSVM typically has high classification accuracy for linearly separable data. However, for nonlinearly separable data, it usually has poor performance. For this type of data, the Support Vector Selection and Adaptation (SVSA) method was developed, but its classification accuracy is not very high for linearly separable data in comparison to LSVM. In this paper, we present a new classifier that combines the LSVM with the SVSA, to be called the Hybrid SVM and SVSA method (HSVSA), for classification of both linearly and nonlinearly separable data and remote sensing images as well. The experimental results show that the HSVSA has higher classification accuracy than the traditional LSVM, the nonlinear SVM (NSVM) with the radial basis kernel, and the previous SVSA.

In this paper, we propose a new machine learning algorithm that we named Support Vector Selection and Adaptation (SVSA). Our aim is to achieve the classification performance of the nonlinear support vector machines (SVM) by using only the support vectors of the linear SVM. The proposed method does not require any type of kernels, and requires less computation time compared to the nonlinear SVM. The SVSA algorithm has two steps: selection and adaptation. In the first step, some of the support vectors obtained from linear SVM are selected. Then the selected support vectors are adapted iteratively in the traning algorithm. The proposed method are compared against the linear and nonlinear SVM on sythetic and real remote sensing data. The results show that the proposed SVSA algorithm achieves very close performance to nonlinear SVM without any kernels in less computation time.