SAR time series are powerful assets for forest monitoring. In recent years, they were involved in various classical forest applications such as forest mapping [9] . These applications largely benefited from the advances of Deep Learning, particularly Unsupervised Learning, using Convolutional Autoencoders in applications such as wildfire detection [4] . Not only did purely temporal approaches offer high prediction performance compared to spatiotemporal variants, but the unsupervised autoencoder rivaled its supervised counterparts. The monitoring of forests also involves the detection of under-canopy targets, which could disturb protected environments. The literature mostly relies on classical SAR approaches PolSAR change detection [8] . A recent shift towards the usage of SAR time series displayed promising performance [10] . Thus, to fully exploit the potential of multi-temporal SAR imagery, this paper proposes the usage of unsupervised Deep Learning, particularly the Convolutional Autoencoder, to detect under forest cover objects.

In this work, we investigate the challenges of implementing a large-scale rainforest mapping with Sentinel-1 short time series. A frequent and accurate monitoring of these ecosystems is of utmost importance in the context of environmental policy-making. In particular, we propose to combine additional descriptive features with deep learning to mitigate the effect of seasonal components on SAR backscatter and interferometric coherences over a year of acquisitions in the Amazon forest. Preliminary analyses suggest that precipitation patterns might play a key role in how discernible land cover classes are with respect to the radar-based input data. Moreover, our findings show that such effects may vary in different regions of the rainforest, so that different configurations of ancillary features might be necessary to achieve a large-scale model able to generalize in both seasonal and regional dimensions.

The C-Band radiometric signature of boreal forests is highly seasonal, with apparent correlations to temperature changes. Within these seasonal components, we assume that information related to tree height can be extracted. We apply a one-dimensional Convolutional Neural Network to assess this assumption, intending to retrieve tree height measured by Airborne Laser Scanning from C-Band Sentinel-1 time series. A study site in the Parc National des Grands Jardins, in Québec, Canada, was selected for this analysis. Prediction-wise, we reach an R2 score of 0.45 and an RMSE of 1.84m, following a 4-fold cross-validation, which exhibits a non-negligible influence of the tree height parameter on boreal forest radiometric response in C-Band Synthetic Aperture Radar, despite the presumed fast saturation of this wavelength, when observing forested environments. In addition to performance metrics, we use a gradient-based explainability tool to diagnose the most contributing periods of the input time series to predict tree height to better correlate the seasonal conditions of this parameter’s influence on the forests’ radiometry.

Temporal and spatial information is exploited in order to improve road detection performance in a forested area. Speckle is filtered using an FBR-based SAR time-series speckle filter, thus preserving temporal information. The additional spatial information is provided by the ascending and descending directions of acquisition of the satellite. Instead of a detection being made if the pixel value is lower than the threshold, we require the threshold to be met for every date in the time-series, and from both directions at once. This corresponds to both a temporal and spatial (generally East-West) stability requirements, and tends to reduce false positives as forest is not a stable environment. The technique was applied on a Sentinel-1 dataset over an area in northern Belize.

In this paper, we propose an extension to an existing permittivity inversion scheme by taking into account double bounce scattering from a rotated dihedral geometry. By considering the rotational aspect, we aim to broaden the applicability of the permittivity inversion scheme for rotated double bounce signatures present in Synthetic Aperture Radar (SAR) imagery. The validity is first demonstrated by invoking to the characteristics of the co-pol ratio through simulations. The inversion scheme is then tested on real airborne SAR data to further validate its effectiveness in real-world scenarios.

This paper presents a simplified scheme for complex permittivity inversion of a double bounce scattering using polarimetric measurements. A forward scattering model is employed to establish the relationship between the observed far field and the underlying permittivity profiles. An inversion scheme based on an approximation is then developed to separately obtain the complex permittivity values corresponding to the two surfaces that constitute a dihedral structure. Electromagnetic simulations demonstrate the effectiveness of the proposed method in accurately recovering the complex permittivity values. Using real airborne data provided by ONERA, the proposed scheme was then utilized to test the possibility of extracting the complex permittivity profiles from double bounce scattering signatures in PolSAR imagery. The obtained complex permittivity profiles exhibit a correlation with previous studies, indicating the reliability of the inversion scheme to study the inverse double bounce scattering problem.

Much work has been dedicated to estimating and optimizing workloads in high-performance computing (HPC) and deep learning. However, researchers have typically relied on few metrics to assess the efficiency of those techniques. Most notably, the accuracy, the loss of the prediction, and the computational time with regard to GPUs or/and CPUs characteristics. It is rare to see figures for power consumption, partly due to the difficulty of obtaining accurate power readings. In this paper, we introduce a composite score that aims to characterize the trade-off between accuracy and power consumption measured during the inference of neural networks. For this purpose, we present a new open-source tool allowing researchers to consider more metrics: granular power consumption, but also RAM/CPU/GPU utilization, as well as storage, and network input/output (I/O). To our best knowledge, it is the first fit test for neural architectures on hardware architectures. This is made possible thanks to reproducible power efficiency measurements. We applied this procedure to state-of-the-art neural network architectures on miscellaneous hardware. One of the main applications and novelties is the measurement of algorithmic power efficiency. The objective is to allow researchers to grasp their algorithms' efficiencies better. This methodology was developed to explore trade-offs between energy usage and accuracy in neural networks. It is also useful when fitting hardware for a specific task or to compare two architectures more accurately, with architecture exploration in mind.

This paper introduces a tool for explaining the latent space generated by applying convolutional autoencoders to satellite image time series, entitled Grad-SLAM. We rely on backpropagated gradient interpretation combined with network activation localization. We use the proposed formula for multiple layers of the encoder, then scale and merge the results to generate a single date contribution metric for the generation of the latent space. We illustrate the potential of this method with the study of the unsupervised classification of agricultural Sentinel-1 time series. We show that critical characterizing dates for unsupervised retrieval of a given class are conditioned by the crop type’s radiometric signature and class count. We also present how Grad-SLAM can be used to enhance the understanding of unsupervised classification confusion.

International audience

Methods using data from different sensors are increasingly used, with the rise of machine learning especially. This article presents a technique based on differential entropy, a concept in information theory that measures the randomness of a distribution. This method can take input data from different sensors. For example, by combining SAR and optical images, one can hope to combine the quality of optical images with the availability of SAR images, capable of seeing at night and through clouds. Examples of this technique using both optical and SAR images are presented in this article, showing the advantages and drawbacks of this approach. The different points to be studied will then be discussed.

It is well known in SAR imagery that the materials of the objects illuminated by a transmitter have a substantial impact on the signal collected by the receiver. It is also widely admitted that this effect depends on the polarization of both emitted and received electromagnetic fields. However, this impact has been poorly investigated, probably because of the lack of ground truth information. We propose in this article to summarize our recent findings on how the materials affect the radar response and how we can use this, both for better understanding the radar phenomenology and for permittivity retrieval.

We extend a technique to retrieve effective permittivities developed for real aperture radar to be used in a Synthetic Aperture Radar (SAR). It can remotely retrieve complex permittivities using polarimetric radar signals from a double bounce. We include the technical hurdles in the extension and then present the first results of a test using airborne X-band data obtained by ONERA.

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.

Temporal Convolutional AutoEncoders are used as feature extractors to project time series onto a latent space where similarity detection can be easily performed. This model can generate accurate descriptors of the temporal profile of the input time-series. We apply this algorithm to PolSAR S1 uncoherent SAR time series where the model learns highly discriminative data representations. This reduction method is compared to others such as PCA or Temporal Averaging and is shown to outperform them when leveraging the learnt representation using K-Means clustering.

The spatial and temporal diversity provided by Sentinel-1 SAR images is used to detect archaeological structures. The temporal mean over a year suppresses speckle without reducing spatial resolution. The combined use of the ascending and descending orbits makes it possible to highlight man-made features.

Road detection is a well-known subject in optics. For example, it is now possible to perform a road extraction and calculate a travel time using multiple look angles optic data [1]. In SAR, road extraction is not new neither [2], [3], even if it is less developed. When it comes to detection in SAR images, it is most of the time about detecting a small target with high intensity, and if the data set is temporal, the target is punctual and present at few dates. However, it is different here for roads, as they can be thin but are often long, have low radiometry, and are permanent in time. This article shows that it might be possible to characterize roads compared to their environment using temporal methods. A proposition for road extraction based on these road's characterizations is shown, and ideas about improving the extraction are presented.

Target detection is an important part of research in SAR. However, the detection of small targets has always been complicated as most of the methods are spatial and based on the use of information from neighbouring pixels to characterise more precisely the studied pixel. In addition, data are often filtered to reduce the noise which results in an additional loss of spatial information. In the frame of small targets detection, this loss has to be avoided and to do so we propose to use temporal methods. Instead of using neighbouring pixels, one can use the different values of the pixels over time which allow not to wipe out small targets. The aim of this preliminary study is to show the advantages of temporal methods over spatial methods and to highlight the points to be further investigated in temporal methods. Promising results on intensity and entropy have been obtained.

Man-made structures, such as buildings, bridges, dams, etc. can be considered as quite constant along time, regarding their position and orientation of course but also for their size and composition. Man-made targets, such as vehicles, exhibit also some constant features, as dimensions and materials. However their position and orientation may have changed between two acquisitions. We propose in this study to investigate the potential benefit of time series of polarimetric SAR images for detection of man-made structures and targets. Two strategies are proposed and tested for these cases.

This article illustrates the benefit of crosspolarization, both in intensity and in phase, to discriminate urban areas and vegetation when there is an ambiguity while using the Pauli decomposition. This ambiguity occurs when buildings are rotated with respect to the sensor trajectory. In this case, we show that because the dissymetry is not of the same nature (deterministic vs quasi-random), it is then possible to discriminate these two environments.

Wild fire is an inflexible threat for the environment equilibrium and human lives, consequently it is vital to study its behaviour for the collective interest and scientific knowledge. Satellite and airbone SAR represents an interesting alternative to optical sensors when the conditions are not favorable. For instance, with presence of clouds, heavy rain or smoke the efficiency of optical sensors decrease drastically for imagery purpose. Since SAR sensors overcome these issues, it is interesting to evaluate their capability to detect fire burn-scars as well as active fires. The NASA Jet Propulsion Lab provide airbone multitemporal L-band full polarimetric data for several areas in USA and South America, in particular, some data are available in California where countrysides are know to be affected by regular significant fires. Depending on the severity of the fire and the vegetation, different physical mechanisms might be affected. The aim of this study is to evaluate the benefit of Pauli basis decomposition in the frame of Change Detection algorithm for burn scar detection.

Confusion between the polarimetric radar signatures of rotated buildings and vegetation has been widely studied to compensate for the effect of rotation. We propose in this study to have a new look on this problem. To do this, we have selected different areas in the San Francisco bay and in New Orleans. These selected zones present different orientations with respect to the sensor illuminating the scene, RADARSAT-2 (C-band) for San Francisco or UAVSAR (L-band) for New Orleans. The variations with the orientation angle of the VV/HH and HV/HH responses collected over these areas are almost identical whatever the sensor. These two quantities first grow and rapidly reach a plateau, at a level which is similar to the responses of forests we collected throughout literature. Actually, it seems that, when the rotation angles grows, the polarimetric radar responses of urban areas tend to limit values which are similar to the radar responses of forests.

Deploying neural networks models over embedded devices have an increased interest and many works is ongoing on that topic. Energy consumption, model sizes and inference time are critical issues as explained in the literature. In the context of IoT and edge computing, tradeoff have been studied in order to get a low cost but rapid answer, robust to connection issue exploiting early exiting or distributing deep neural networks. Those approaches exploits the cloud as an endpoint, balancing the load with respect to different computing capabilities. In this paper, we propose to extend those approaches to networks of embedded devices such as a swarm of drones, where every device has the same computing capabilities (in terms of energy and speed). Computing load may be balanced among the whole swarm in order to maximise either the lifetime of specific devices or lifetime of the whole swarm. We develop criteria to best cut and distribute those networks, validate them through power measurement and express the different tradeoffs we have to address.

In this paper the development of a new method to extract the two complex relative permittivities of a dihedral structure composed of two surfaces in monostatic configuration is exposed. Provided a dominant double bounce mechanism, we only need in theory the copolarised far-field measurements HH and VV at one incidence angle. As the permittivity of natural material is strongly dependent on its moisture content, the method is used here to infer the water content of the soil. To do so we worked with simulated data using the Geometric Optic method implemented in FEKO. First results show that some incidence angle range has to be preferred to accurately determine the permittivity of the horizontal surface.

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.

The double bounce mechanism is the main characteristic in urban areas. This paper describes two physical phenomena that can decrease it (and even suppress it) and thus alter the radar analysis of cities.

When the incidence angle is equal or about the Brewster's angle, a lossy medium favors the transmission of VV polarization. As a result, there may be a strong difference between the co-polarization components of the reflected waves. For a dihedral-type scattering mechanism, this effect is enhanced and occurs for a wider range of incidence angles. This situation - typically present in urban areas - may induce significant changes in the polarimetric signatures.

The use of multiplicative beamforming is described to improve the azimuthal resolution of existing high-frequency surface-wave radar (HFSWR). An existing method by Davies and Ward is used to improve the resolution by a factor of more than two for HFSWR with 16 antennas. The long integration time of HFSWR suppresses the cross products which are the usual drawbacks of this process. The concept is then validated on experimental data with synthetic targets.

As a long term objective we would like to define the required conditions to detect oil spills, using High Frequency Surface Wave Radar (HFSWR). Assuming that the presence of oil spills on the sea surface modifies the surface tension which in turn affects the dynamics of the sea, we would like to determine the minimal surface tension which makes this phenomenon observable on a Doppler spectrum. For this purpose we have developed a simulator based on a realistic modeling of the sea and of its temporal variation as well as on an exact modeling of the interactions between the electromagnetic waves and the environment. We have chosen to implement a full wave model called ELSEM3D, developed by ONERA. This tool is based on the Method of Moments (MoM) which can be accelerated by the Fast Multipole Method (FMM). The issues of such a modeling are numerous: computational time due to the number of temporal realizations required to generate a Doppler spectrum and the dimension of the scene. In this paper, we will discuss the issues raised by this study and outline how we can overcome them. We will also present some simulated Doppler spectra for different sea states and radar configurations.