Building segmentation from remote sensing data has been boosted in recent years by the advances in data processing algorithms and the improved ability of sensors to capture very high spatial resolution (VHSR) images. The prevailing approaches to this task are built upon deep semantic segmentation networks learned on ortho geometry, i.e., orthorectified images. In this study, we propose to deal with the building segmentation using multiview Pléiades satellite images at the perfect sensor (PS) geometry instead of the ortho geometry. This has two main advantages: 1) it frees the segmentation workflow from geometric imprecision that may arise after the image rectification step and 2) it allows the physical augmentations of the ground truth (GT) by reprojecting building objects on each of the multiview acquisitions. The GT reprojection process makes use of rational polynomial coefficients (RPCs) provided as image metadata and a fine scale digital surface model (DSM). We assess the benefit of our proposal using a U-Net encoder-decoder learned on a dataset composed of tri-stereo Pléiades acquisitions over six French cities. Experimental results demonstrate the significance of the proposal especially an enhanced generalization capability for the building segmentation.

Continual learning in dynamic environments is a challenge for large-scale machine learning models. This research addresses Domain Incremental Learning (DIL), a setting where the goal is to incrementally increase the input data scope of a model. More specifically, we investigate the possibility of using foundation models (FMs) as a fixed feature extractor combined with a PPCA that can be sequentially and accurately updated. Focusing on the classification of VHR remote sensing (RS) images, we show on the FLAIR#1 dataset that this simple DIL strategy achieves competitive accuracy compared to memory-based baselines across different pre-trained sources. We also compare different types of foundation models and highlight the importance of data diversity over data specialization to improve the quality of FMs.

Nowadays, the state-of-the-art in building extraction from remote sensing data is largely driven by deep learning methods. While it is feasible to supervise a model to output in an end-to-end fashion the vector footprints of building instances from high spatial resolution images, the availability of ground truth with high quality annotations may limit the success of this approach, which in fact is highly beneficial for downstream applications. An alternative strategy might be the post processing of building segmentation masks using geometric priors to obtain sharp and regularized polygonal footprints. Nonetheless, misalignment frequently occurs after the regularization, typically in suburban areas, since each building instance is processed independently of those in its neighborhood. To address misalignment, we propose in this study to take into account neighborhood through Voronoï partitions and identify groups of buildings that share similar geometric properties. The buildings constituting a local group are finally realigned and regularized as well using the same edge orientations extracted with the Line Segment Detector algorithm. Our proposal is assessed on a Pléiades image depicting a residential area of Montpellier in southern France.

3D Geospatial information plays a key role in many soaring sectors such as sustainable and smart cities, climate monitoring, ecological mobility, and economic intelligence. The availability of huge volumes of satellite, airborne and insitu data now makes this production feasible at large scale. It needs nonetheless a certain level of manual intervention to secure the level of quality, which prevents mass production. This paper presents the AI4GEO program that aims at developing an end to end solution to produce automatically qualified 3D Digital model at scale together with multiple layers of information.

Despite the intense development of deep neural networks for computer vision, and especially semantic segmentation, their application to Earth Observation data remains usually below accuracy requirements brought by real-life scenarios. Even if well-known deep learning methods produce excellent results, they tend to be over-confident and cannot assess how relevant their predictions are. In this work, a Bayesian deep learning method, based on Monte Carlo Dropout, is proposed to tackle semantic segmentation of aerial and satellite images. Bayesian deep learning can provide both a semantic segmentation and uncertainty maps. Based on the popular U-Net architecture, our model achieves semantic segmentation with high accuracy, e.g. F1-score and overall accuracy respectively reaching 90.84% and 93.22% on a public standard dataset. Uncertainty maps, also derived from our model, show a strong interest in qualitative evaluation of the segmentation and in the improvement of the database.

According to the United Nations, 70% of earth population is going to live in cities by 2050. Given this fast urban evolution, urban monitoring is a key process to qualify sustainable development. Vertical changes need to be assessed, and various methods for 3D change detection have been published. However, there is no common quantitative benchmark assessing their performance in urban areas yet. In this paper, we aim to fill this gap and introduce a simulation tool to generate synthetic 3D point cloud data in a well-controlled scenario. These data are then used to compare qualitatively and quantitatively representative 3D change detection methods for urban areas. These methods are based on distance computation (DSMd, C2C, M3C2), traditional machine learning (RF with stability feature) and deep learning (Feed Forward and Siamese networks). We distinguish between binary and multi-class classification of changes at different levels (3D points, 2D pixels, and 2D patches). While deep neural networks have led to numerous success in remote sensing, we show that they do not systematically outperform more simple methods for 3D change detection. Besides, the existing networks are limited to 2D patches while outputs at the pixel or point scale are more attractive.

Detection of new infrastructures (commercial, logistics, industrial or residential) from satellite images constitutes a proven method to investigate and follow economic and urban growth. The level of activities or exploitation of these sites may be hardly determined by building inspection, but could be inferred from vehicle presence from nearby streets and parking lots. We present in this paper two deep learning-based models for vehicle counting from optical satellite images coming from the Pleiades sensor at 50-cm spatial resolution. Both segmentation (Tiramisu) and detection (YOLO, You Only Look Once) architectures were investigated. These networks were adapted, trained and validated on a data set including 87k vehicles, annotated using an interactive semi-automatic tool developed by the authors. Experimental results show that both segmentation and detection models could achieve a precision rate higher than 85 % with a recall rate also high (76.4 % and 71.9 % for Tiramisu and YOLO respectively).

In the field of remote sensing and more specifically in Earth Observation, new data are available every day, coming from different sensors. Leveraging on those data in classification tasks comes at the price of intense labelling tasks that are not realistic in operational settings. While domain adaptation could be useful to counterbalance this problem, most of the usual methods assume that the data to adapt are comparable (they belong to the same metric space), which is not the case when multiple sensors are at stake. Heterogeneous domain adaptation methods are a particular solution to this problem. We present a novel method to deal with such cases, based on a modified cycleGAN version that incorporates classification losses and a metric space alignment term. We demonstrate its power on a land use classification tasks, with images from both Google Earth and Sentinel-2.

Through various initiatives, CNES, the French space agency, has been involved in major disaster and environment monitoring from Space for many years and in particular in the International Charter which delivers satellite data to nations affected by major disasters and in the THEIA organization which promotes the use of satellite data to monitor human and climate impacts on environment.Thus, CNES developed in 2014 a reference framework to generate value-added products from satellite data. Since 2017, considering the increase of data and processing requests, CNES has decided to move to a new framework based on big data and cloud technologies and extended to data analytics.This paper will first introduce THEIA and Charter initiatives. It will present the current framework in operation and its limitations. It will then focus on the innovative approach to handle big data and analytics needs and finally presents the first results and perspectives.

Processing large very high-resolution remote sensing images on resource-constrained devices is a challenging task because of the large size of these data sets. For applications such as environmental monitoring or natural resources management, complex algorithms have to be used to extract information from the images. The memory required to store the images and the data structures of such algorithms may be very high (hundreds of gigabytes) and therefore leads to unfeasibility on commonly available computers. Segmentation algorithms constitute an essential step for the extraction of objects of interest in a scene and will be the topic of the investigation in this paper. The objective of the present work is to adapt image segmentation algorithms for large amounts of data. To overcome the memory issue, large images are usually divided into smaller image tiles, which are processed independently. Region-merging algorithms do not cope well with image tiling since artifacts are present on the tile edges in the final result due to the incoherencies of the regions across the tiles. In this paper, we propose a scalable tile-based framework for region-merging algorithms to segment large images, while ensuring identical results, with respect to processing the whole image at once. We introduce the original concept of the stability margin for a tile. It allows ensuring identical results to those obtained if the whole image had been segmented without tiling. Finally, we discuss the benefits of this framework and demonstrate the scalability of this approach by applying it to real large images.

Large scale segmentation remains a challenging task because of time and memory consuming. A usual strategy to process efficiently a large volume of data is to divide into chunks to be processed separately, either sequentially to reduce memory footprint or in parallel in order to speed up the computation. In image processing in general this boils down to dividing the input image into tiles. However, for image segmentation, the tile splitting usually leads incoherent segments on the borders of the tiles even when some overlap between the tiles is applied. In this paper we propose a new strategy making possible the tiling for image segmentation algorithms while maintaining the accuracy of the final results. Specifically, we focus on iterative region merging methods but the strategy can be extended to any segmentation algorithm. The introduction of the local mutual best fitting concept and the area of influence of a segment allows to establish a new methodology of segmentation based on three phases: the tile-based reduction, the iterative reduction and the completion of the segmentation. This new methodology was applied on a large Pleiades HR image with success proving the feasibility of the approach.