In this paper, we propose a remote-sensing scene-classification method based on vision transformers. These types of networks, which are now recognized as state-of-the-art models in natural language processing, do not rely on convolution layers as in standard convolutional neural networks (CNNs). Instead, they use multihead attention mechanisms as the main building block to derive long-range contextual relation between pixels in images. In a first step, the images under analysis are divided into patches, then converted to sequence by flattening and embedding. To keep information about the position, embedding position is added to these patches. Then, the resulting sequence is fed to several multihead attention layers for generating the final representation. At the classification stage, the first token sequence is fed to a softmax classification layer. To boost the classification performance, we explore several data augmentation strategies to generate additional data for training. Moreover, we show experimentally that we can compress the network by pruning half of the layers while keeping competing classification accuracies. Experimental results conducted on different remote-sensing image datasets demonstrate the promising capability of the model compared to state-of-the-art methods. Specifically, Vision Transformer obtains an average classification accuracy of 98.49%, 95.86%, 95.56% and 93.83% on Merced, AID, Optimal31 and NWPU datasets, respectively. While the compressed version obtained by removing half of the multihead attention layers yields 97.90%, 94.27%, 95.30% and 93.05%, respectively.

We present a novel automatic and unsupervised change-detection approach specifically oriented to the analysis of multitemporal single-channel single-polarization synthetic aperture radar (SAR) images. This approach is based on a closed-loop process made up of three main steps: (1) a novel preprocessing based on a controlled adaptive iterative filtering; (2) a comparison between multitemporal images carried out according to a standard log-ratio operator; and (3) a novel approach to the automatic analysis of the log-ratio image for generating the change-detection map. The first step aims at reducing the speckle noise in a controlled way in order to maximize the discrimination capability between changed and unchanged classes. In the second step, the two filtered multitemporal images are compared to generate a log-ratio image that contains explicit information on changed areas. The third step produces the change-detection map according to a thresholding procedure based on a reformulation of the Kittler-Illingworth (KI) threshold selection criterion. In particular, the modified KI criterion is derived under the generalized Gaussian assumption for modeling the distributions of changed and unchanged classes. This parametric model was chosen because it is capable of better fitting the conditional densities of classes in the log-ratio image. In order to control the filtering step and, accordingly, the effects of the filtering process on change-detection accuracy, we propose to identify automatically the optimal number of despeckling filter iterations [Step 1] by analyzing the behavior of the modified KI criterion. This results in a completely automatic and self-consistent change-detection approach that avoids the use of empirical methods for the selection of the best number of filtering iterations. Experiments carried out on two sets of multitemporal images (characterized by different levels of speckle noise) acquired by the European Remote Sensing 2 satellite SAR sensor confirm the effectiveness of the proposed unsupervised approach, which results in change-detection accuracies very similar to those that can be achieved by a manual supervised thresholding.

The aim of this paper is twofold. First, we present a thorough experimental study to show the superiority of the generalization capability of the support vector machine (SVM) approach in the automatic classification of electrocardiogram (ECG) beats. Second, we propose a novel classification system based on particle swarm optimization (PSO) to improve the generalization performance of the SVM classifier. For this purpose, we have optimized the SVM classifier design by searching for the best value of the parameters that tune its discriminant function, and upstream by looking for the best subset of features that feed the classifier. The experiments were conducted on the basis of ECG data from the Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) arrhythmia database to classify five kinds of abnormal waveforms and normal beats. In particular, they were organized so as to test the sensitivity of the SVM classifier and that of two reference classifiers used for comparison, i.e., the k-nearest neighbor (kNN) classifier and the radial basis function (RBF) neural network classifier, with respect to the curse of dimensionality and the number of available training beats. The obtained results clearly confirm the superiority of the SVM approach as compared to traditional classifiers, and suggest that further substantial improvements in terms of classification accuracy can be achieved by the proposed PSO-SVM classification system. On an average, over three experiments making use of a different total number of training beats (250, 500, and 750, respectively), the PSO-SVM yielded an overall accuracy of 89.72% on 40438 test beats selected from 20 patient records against 85.98%, 83.70%, and 82.34% for the SVM, the kNN, and the RBF classifiers, respectively.

Recent remote sensing literature has shown that support vector machine (SVM) methods generally outperform traditional statistical and neural methods in classification problems involving hyperspectral images. However, there are still open issues that, if suitably addressed, could allow further improvement of their performances in terms of classification accuracy. Two especially critical issues are: 1) the determination of the most appropriate feature subspace where to carry out the classification task and 2) model selection. In this paper, these two issues are addressed through a classification system that optimizes the SVM classifier accuracy for this kind of imagery. This system is based on a genetic optimization framework formulated in such a way as to detect the best discriminative features without requiring the a priori setting of their number by the user and to estimate the best SVM parameters (i.e., regularization and kernel parameters) in a completely automatic way. For these purposes, it exploits fitness criteria intrinsically related to the generalization capabilities of SVM classifiers. In particular, two criteria are explored, namely: 1) the simple support vector count and 2) the radius margin bound. The effectiveness of the proposed classification system in general and of these two criteria in particular is assessed both by simulated and real experiments. In addition, a comparison with classification approaches based on three different feature selection methods is reported, i.e., the steepest ascent (SA) algorithm and two other methods explicitly developed for SVM classifiers, namely: 1) the recursive feature elimination technique and 2) the radius margin bound minimization method

Scene classification is a highly useful task in Remote Sensing (RS) applications. Many efforts have been made to improve the accuracy of RS scene classification. Scene classification is a challenging problem, especially for large datasets with tens of thousands of images with a large number of classes and taken under different circumstances. One problem that is observed in scene classification is the fact that for a given scene, only one part of it indicates which class it belongs to, whereas the other parts are either irrelevant or they actually tend to belong to another class. To address this issue, this paper proposes a deep attention Convolutional Neural Network (CNN) for scene classification in remote sensing. CNN models use successive convolutional layers to learn feature maps from larger and larger regions (or receptive fields) of the scene. The attention mechanism computes a new feature map as a weighted average of these original feature maps. In particular, we propose a solution, named EfficientNet-B3-Attn-2, based on the pre-trained EfficientNet-B3 CNN enhanced with an attention mechanism. A dedicated branch is added to layer 262 of the network, to compute the required weights. These weights are learned automatically by training the whole CNN model end-to-end using the backpropagation algorithm. In this way, the network learns to emphasize important regions of the scene and suppress the regions that are irrelevant to the classification. We tested the proposed EfficientNet-B3-Attn-2 on six popular remote sensing datasets, namely UC Merced, KSA, OPTIMAL-31, RSSCN7, WHU-RS19, and AID datasets, showing its strong capabilities in classifying RS scenes.