The classification of urban environments poses significant challenges for polarimetric synthetic aperture radar (SAR), since rotated urban blocks violate the assumption of reflection symmetry held by many decomposition approaches. Vast sections of cities appear to be dominated by volume scattering due to the introduction of coherent depolarization. Attempts to model this slanted double bounce must contend with the fact that dihedrals rotated about the vertical axis do not support significant backscatter and require a rough ground surface to affect a return path. In this paper, the concept of an effective dihedral is introduced, with one plate coincident with the building wall and one plate associated with some ground facet, oriented so as to support double bounce. The rotation of this effective dihedral can be readily related to the polarization orientation angle and used to quantify co-pol and cross-pol phases. Theoretical results are confirmed with L-band and C-band polarimetric SAR imagery of San Francisco, as well as by broadband laboratory experiments with modeled buildings. In addition, the utility of using cross-pol phase to classify urban regions in the Los Angeles basin is demonstrated with ascending and descending dual-pol Radarsat data.

Buildings that are rotated with respect to the sensor trajectory could be erroneously classified as vegetated areas in the Pauli basis, and subsequently in many decomposition theorems despite the considerable amount of work done to solve that issue. This misjudgement is linked to the high level of their cross-polarized contribution. Using electromagnetic simulation tools and image analysis, we study the value of these cross-polarization components. We show that forested areas and cities exhibit significantly different cross-polarization levels; indeed, the origin of these components is actually distinct. Based on that, to discriminate between the two environments, we introduce an extension to the Pauli basis where the cross polarization is split into two classes, one for rotated dihedrals and the other for random scatterers. This approach is then tested on two synthetic aperture radar images: the first acquired at C-band using RADARSAT-2 over Downtown San Francisco and the second using RAMSES at X-band over an industrial area near Paris.

With an increase in the amount of natural disasters, the combined use of cloud-penetrating Synthetic Aperture Radar and deep learning becomes unavoidable for their monitoring. This article proposes a methodology for forest fire detection using unsupervised location-expert autoencoders and Sentinel-1 SAR time series. The models are trained on SAR multitemporal images over a specific area using a reference period and extract any deviating time series over that same area for the test period. We present three variations of the autoencoder, incorporating either temporal features or spatiotemporal features, and we compare it against a state-of-the-art supervised autoencoder. Despite their limitations, we show that unsupervised approaches are on par with supervised techniques, performance-wise. A specific architecture, the fully temporal autoencoder, stands out as the best-performing unsupervised approach by leveraging temporal information of Sentinel-1 time series using one-dimensional convolutional layers. The approach is generic and can be applied to many applications, though we focus here on forest fire detection in Canadian boreal forests as a successful use case.

Polarimetric Synthetic Aperture Radar (PolSAR) images containing cities may exhibit misclassified areas when using polarimetric decompositions. Several articles relate this problem to the effects of orientation between the facades of buildings and the acquisition trajectory. Materials also play a role in polarimetric behavior. This paper deals with this combined effect of material and orientation. It analyzes different sets of data, airborne or space-borne, at L-, C- and X-bands, and for different orientation angles. It shows that considering dielectric dihedral rather than metallic in the polarimetric mechanism of double-bounce has a very important impact on the differences of intensities between the channels HH and VV. This difference is very important for small angles of orientation, and then decreases for large angles. Furthermore, the curves of the ratios between polarimetric intensities as a function of the orientation angle vary little with the materials and the frequencies encountered in all the scenarios envisaged. The signal of the ratio VV/HH raises a plateau around −1 dB for orientations higher than 30°. We also observe a plateau for HV/HH, but with a value around −5 dB.

We present a fully unsupervised learning pipeline, which involves both a projection method and a clustering algorithm dedicated to the pixel-wise classification of multitemporal SAR images. We design a Convolutional Autoencoder as the method to project our time series onto a lower dimensional latent space, where semantically similar temporal signals are placed close together. The additional use of convolutional layers as feature extraction steps allows us to exploit the sequential nature of time series, exhibiting higher representation performance than fully connected layers. The extracted clusters can encapture different semantic levels to either separate classes or extract outlying temporal signals. The application of this method to crop-types mapping enables the extraction of major crop-types within a scene, without supervision. In a labeled context, this method also allows for the extraction of outlying profiles which can lead to the discovery of mislabeled time series.

Change detection (CD) in SAR (Synthethic Aperture Radar) images has been widely studied in recent years and has become increasingly attractive due to the growth of available datasets. The potential of CD has been shown in different fields, including disaster monitoring and military applications. Access to multi-temporal SAR images of the same scene is now possible, and therefore we can improve the performance and the interpretation of CD. Apart from specific SAR campaign measurements, the ground truth of the scene is usually unknown or only partially known when dealing with open data. This is a critical issue when the purpose is to detect targets, such as vehicles or ships. Indeed, typical change detection methods can only provide relative changes; the actual number of targets on each day cannot be determined. Ideally, this change detection should occur between a target-free image and one with the objects of interest. To do so, we propose to benefit from pixels’ intrinsic temporal behavior to compute a frozen background reference (FBR) image and perform change detection from this reference image. We will then consider that the scene consists only of immobile objects (e.g., buildings and trees) and removable objects that can appear and disappear from acquisition to another (e.g., cars and ships). Our FBR images will, therefore, aim to estimate the immobile background of the scene to obtain, after change detection, the exact amount of targets present on each day. This study was conducted first with simulated SAR data for different number of acquisition dates and Signal-to-Noise Ratio (SNR). We presented an application in the region of Singapore to estimate the number of ships in the study area for each acquisition.