Feature learning algorithms that use deep neural networks have shown to outperform traditional hand-crafted feature extraction methods when applied to satellite image time series. This learned features can be used in a multitude of applications such as classification, semantic segmentation, and change detection, among others. In this paper, we employ a feature learning technique to extract representative features from multimission polarimetric SAR (PolSAR) satellite image time series (SITS). To this end, we implement a formulation combining a 1-dimensional convolutional neural network and a stacked auto encoder. We performed experiments on a multimission and multitemporal PolSAR dataset and validated the extracted features through their utility as features in an unsupervised classification problem.

Decision fusion allows making a common decision by combining multiple opinions. In the context of remote sensing classification, such techniques are of great importance in all the cases where data collected by multiple sensors are merged into a final decision. Decision fusion may be used to combine the posterior probabilities associated with the output of single classifiers when applied to single sensor data. Meanwhile, techniques such as Markov Random Fields (MRFs) can integrate contextual information in the fusion process and are commonly used in classification. However, in the context of very large scale mapping (e.g., for global climate change monitoring), computation time can be critical and the application of both data fusion and spatial-contextual modeling comes with several constraints. In this paper, we propose a Bayesian decision fusion approach for optical-SAR image classification, integrated with a fast formulation of the iterated conditional modes (ICM) MRF-optimization algorithm based on a convolution operation. he validation on wide areas of Siberia proved the scalability and efficiency of the method for large scale applications.

Scene classification of remote sensing images is a challenging task due to the complexity and variety of natural scenes. In recent years, Convolutional Neural Networks (CNNs) have achieved impressive performances in many remote sensing scene classification benchmarks. However, in CNNs the long-range visual dependencies are often neglected due to the local filter design, leading to suboptimal performances in cluttered scenes such as urban areas. Recently proposed Transformer architecture resolved this issue by taking a broader neighborhood into account through the multi-head self-attention component. In this paper, we propose a novel method which borrows ideas from “knowledge distillation” and applied to recent vision Transformers. Specifically, we propose a compound loss computed on a Transformer-based student and a CNN teacher in a joint fashion and utilize it for the task of single-label scene classification. Because of the student's capability in capturing long-range visual dependencies, along with the inductive bias inherited from the teacher, our proposed model improves the classification accuracy on four well-known datasets compared to state-of-the-art approaches.

The automatic registration of image pairs composed of optical and synthetic aperture radar (SAR) images is a highly challenging task because of the inherently different physical, statistical, and textural properties of the input data. Information-theoretic measures capable of comparing local intensity distributions are often used for multisensor optical-SAR registration. Moreover, the growing availability of such heterogeneous data from current space missions require multisensor registration methods able to run on large-scale datasets with acceptable computation times. In this paper, a novel method is proposed combining information-theoretic area-based registration with a sequential image tiling strategy. Experiments with optical-SAR data collected by a variety of sensors (Sentinel, Landsat, ERS, etc.) suggest both qualitatively and quantitatively the effectiveness of the proposed strategy in achieving accurate registration with low computational cost.

Convolutional neural networks (CNNs) represent the new reference approach for semantic segmentation of very-high-resolution (VHR) images, due to their ability to automatically capture semantic information while learning relevant features. However, as for most supervised methods, the map accuracy depends on the quantity and quality of ground truth (GT) used to train them. The use of densely annotated data (i.e., a detailed, exhaustive, pixel-level GT) allows to obtain effective CNN models but normally implies high efforts in annotation. Such ground truth is often available in benchmark datasets on which new methods are tested, but not on real data for land-cover applications, where only sparse annotations might be sufficiently cost effective. A CNN model trained with such incomplete GT maps has the tendency to smooth object boundaries because they are never precisely delineated in the GT. To cope with those shortcomings, we propose to exploit the intermediate activation maps of the CNN and to deploy a semisupervised fully connected conditional random field (CRF). In comparison with competitors using the same sparse annotations, the proposed method is able to better fill part of the performance gap compared to a CNN trained on the densely annotated, but generally unavailable, GTs.

Synthetic aperture radar (SAR) and optical satellite image registration is a field that developed in the last decades and gave rise to a great number of approaches. The registration process is composed of several steps: feature definition, feature comparison and optimization of a geometric transformation between the images. Feature definition can be done using simple traditional filtering or more complex deep learning (DL) methods. In this paper, two traditional approaches and a DL approach are compared. One can then wonder if the complexity of DL is worth to address the registration task. The aim of this paper is to quantitatively compare approaches rooted in distinct methodological areas on two common datasets with different resolutions. The comparison suggests that, although more complex, the DL approach is more precise than traditional methods.

The automatic registration of multisensor remote sensing images is a highly challenging task due to the inherently different physical, statistical, and textural properties of the input data. In the present paper, this problem is addressed in the case of optical-SAR images by proposing a novel method based on deep learning and area-based registration concepts. The method integrates a conditional generative adversarial network (cGAN), an area-based cross-correlation-type l2similarity metric, and the COBYLA constrained maximization algorithm. Whereas correlation-type metrics are typically ineffective in the application to multisensor registration, the proposed approach allows exploiting the image translation capabilities of cGAN architectures to enable the use of an l2similarity metric, which favors high computational efficiency. Experiments with Sentinel-1 and Sentinel-2 data suggest the effectiveness of this strategy and the capability of the proposed method to achieve accurate registration.

Convolutional Neural Networks (CNNs) have become the new standard for semantic segmentation of very high resolution images. But as for other methods, the map accuracy depends on the quantity and quality of ground truth used to train them. Having densely annotated data, i.e. a detailed, pixel-level ground truth (GT), allows obtaining effective models, but requires high efforts in annotation. For this reason, it is more common and efficient to work with point or scribbled annotations rather than with dense ones. A CNN model trained with such incomplete ground truths tends to mischaracterize the shapes of the objects and to be inaccurate near their boundaries. We propose to use an approximation of a fully connected Conditional Random Field (CRF) to solve these issues, in which long range connections are accounted for through auxiliary nodes based on clustering of CNN activation features. Experiments on the ISPRS Vaihingen benchmark, where a CNN is trained only with a non-dense, scribbled ground truth, show that the proposed method can fill part of the performance gap with respect to models trained on the densely annotated, but unrealistic, ground truth.