Self-supervised learning (SSL) has revolutionized representation learning in Remote Sensing (RS), advancing Geospatial Foundation Models (GFMs) to leverage vast unlabeled satellite imagery for diverse downstream tasks. Currently, GFMs primarily employ objectives like contrastive learning or masked image modeling, owing to their proven success in learning transferable representations. However, generative diffusion models, which demonstrate the potential to capture multi-grained semantics essential for RS tasks during image generation, remain underexplored for discriminative applications. This prompts the question: can generative diffusion models also excel and serve as GFMs with sufficient discriminative power? In this work, we answer this question with SatDiFuser, a framework that transforms a diffusion-based generative geospatial foundation model into a powerful pretraining tool for discriminative RS. By systematically analyzing multi-stage, noise-dependent diffusion features, we develop three fusion strategies to effectively leverage these diverse representations. Extensive experiments on remote sensing benchmarks show that SatDiFuser outperforms state-of-the-art GFMs, achieving gains of up to +5.7% mIoU in semantic segmentation and +7.9% F1-score in classification, demonstrating the capacity of diffusion-based generative foundation models to rival or exceed discriminative GFMs. The source code is available at: https://github.com/yurujaja/SatDiFuser.

Vessel identification through radar imagery aims to address the complex task of not only detecting, but also classifying ships using SAR images. The goal of the xView3 challenge is to test the capabilities of Sentinel-1 data for the identification of fishing vessels. This initiative plays a crucial role in combating illegal fishing activities and mitigating the associated environmental damage. By fostering advancements in ship classification techniques, the challenge seeks to contribute significantly to global efforts in monitoring and preserving marine ecosystems, particularly by curbing illicit fishing practices.Our paper introduces an end-to-end ship identification method that enhances the challenge-winning CircleNet solution. This improvement is achieved through extending the method for analyzing multi-temporal SAR data, facilitating more robust predictions by effectively filtering out spurious detections of stationary objects.

Floods are one of the most destructive disasters worldwide and although they largely happen in rural, ruther than in urban areas, it is in the urban areas that substantial destruction of infrastructures is observed. Thus, cost effective methods to monitor flood damage and extent are required. In this paper, we investigate the implementation of U-Net on satellite and drone image dataset such as xBD and EDDA for building damage assessment in Mozambique. The recently published dataset EDDA was created by the National Institute for Disaster Management (INGD) and comprises drone imagery of Beira, in Mozambique. Using them, we obtained a dice score of 0.76 on building localization (BL) and mean intersection over the union (mIoU) of 0.54 on damage classification (DC). These are promising results considering that many datasets lack detailed information on African buildings. We also use some pre-trained models models such as ResNet for BL and DC.

Accurate estimation of building heights is essential for urban planning, infrastructure management, and environmental analysis. In this study, we propose a supervised Multimodal Building Height Regression Network (MBHR-Net) for estimating building heights at 10m spatial resolution using Sentinel-1 (S1) and Sentinel-2 (S2) satellite time series. S1 provides Synthetic Aperture Radar (SAR) data that offers valuable information on building structures, while S2 provides multispectral data that is sensitive to different land cover types, vegetation phenology, and building shadows. Our MBHR-Net aims to extract meaningful features from the S1 and S2 images to learn complex spatio-temporal relationships between image patterns and building heights. The model is trained and tested in 10 cities in the Netherlands. Root Mean Squared Error (RMSE), Intersection over Union (IOU), and R-squared (R2) score metrics are used to evaluate the performance of the model. The preliminary results (3.73m RMSE, 0.95 IoU, 0.61 R2) demonstrate the effectiveness of our deep learning model in accurately estimating building heights, showcasing its potential for urban planning, environmental impact analysis, and other related applications.

Change detection using earth observation data plays a vital role in quantifying the impact of disasters in affected areas. While data sources like Sentinel-2 provide rich optical information, they are often hindered by cloud cover, limiting their usage in disaster scenarios. However, leveraging pre-disaster optical data can offer valuable contextual information about the area such as landcover type, vegetation cover, soil types, enabling a better understanding of the disaster’s impact. In this study, we develop a model to assess the contribution of pre-disaster Sentinel-2 data in change detection tasks, focusing on disaster-affected areas. The proposed Context-Aware Change Detection Network (CACDN) utilizes a combination of pre-disaster Sentinel-2 data, pre and post-disaster Sentinel-1 data and ancillary Digital Elevation Models (DEMs) data. The model is validated on flood and landslide detection and evaluated using three metrics: Area Under the Precision-Recall Curve (AUPRC), Intersection over Union (IoU), and mean IoU. The preliminary results show significant improvement (4%, AUPRC, 3-7% IoU, 3-6% mean IoU) in model’s change detection capabilities when incorporated with pre-disaster optical data reflecting the effectiveness of using contextual information for accurate flood and landslide detection.

Above Ground Biomass is an important variable as forests play a crucial role in mitigating climate change as they act as an efficient, natural and cost-effective carbon sink. Traditional field and airborne LiDAR measurements have been proven to provide reliable estimations of forest biomass. Nevertheless, the use of these techniques at a large scale can be challenging and expensive. Satellite data have been widely used as a valuable tool in estimating biomass on a global scale. However, the full potential of dense multi-modal satellite time series data, in combination with modern deep learning approaches, has yet to be fully explored. The aim of the "BioMassters" data challenge and benchmark dataset is to investigate the potential of multi-modal satellite data (Sentinel-1 SAR and Sentinel-2 MSI) to estimate forest biomass at a large scale using the Finnish Forest Centre's open forest and nature airborne LiDAR data as a reference. The performance of the top three baseline models shows the potential of deep learning to produce accurate and higher-resolution biomass maps. Our benchmark dataset is publically available at https://huggingface.co/datasets/nascetti-a/BioMassters (doi:10.57967/hf/1009) and the implementation of the top three winning models are available at https://github.com/drivendataorg/the-biomassters.

In this study, a Semi-Supervised Learning (SSL) method for improved urban change detection from bi-temporal image pairs is presented. The proposed method employs a Dual-Task Siamese Difference network that not only predicts changes with the difference decoder, but also segments buildings for both images with a semantic decoder. First, the architecture was modified to produce a second change prediction derived from the semantic predictions. Second, SSL was used to improve supervised change detection. For unlabeled data, we designed a loss that encourages the network to predict consistent changes across the two change outputs. The proposed method was tested on urban change detection using the SpaceNet7 dataset. SSL achieved improved results compared to three fully supervised benchmarks. Code for the paper is available at https://github.com/SebastianHafner/SiameseSSL.git.

This work presents a Plug- In for the Opticks open source soft-ware implementing an unsupervised workflow for building roof extraction from Light Detection and Ranging (LiDAR) data. In particular, a computer vision approach is employed to segment the points belonging to different objects (buildings, trees, etc.), whereas the RANSAC algorithm, the core of the proposed workflow, is used recursively for identifying the buildings and to model their roofs. The preliminary re-sults, qualitatively evaluated, are encouraging: the proposed roof extraction workflow works generally well - the 80% of the roofs are completely or partially modeled - but shows some issues with buildings characterized by several pitches with low slopes and/or located in proximity of dense vegetation.

Due to climate and land-use change, natural disasters such as flooding have been increasing in recent years. Timely and reliable flood detection and mapping can help emergency response and disaster management. In this work, we propose a flood detection network using bi-temporal SAR acquisitions. The proposed segmentation network has an encoder-decoder architecture with two Siamese encoders for pre and post-flood images. The network's feature maps are fused and enhanced using attention blocks to achieve more accurate detection of the flooded areas. Our proposed network is evaluated on publicly available Sen1Flood11 [1] benchmark dataset. The network outperformed the existing state-of-the-art (uni-temporal) flood detection method by 6% IOU. The experiments highlight that the combination of bi-temporal SAR data with an effective network architecture achieves more accurate flood detection than uni-temporal methods.

Urbanization is progressing rapidly around the world. With sub-weekly revisits at global scale, Sentinel-1 synthetic aperture radar (SAR) and Sentinel-2 multispectral imager (MSI) data can play an important role for monitoring urban sprawl to support sustainable development. In this letter, we proposed an urban change detection (CD) approach featuring a new network architecture for the fusion of SAR and optical data. Specifically, a dual stream concept was introduced to process different data modalities separately, before combining extracted features at a later decision stage. The individual streams are based on U-Net architecture that is one of the most popular fully convolutional networks used for semantic segmentation. The effectiveness of the proposed approach was demonstrated using the Onera Satellite CD (OSCD) dataset. The proposed strategy outperformed other U-Net-based approaches in combination with unimodal data and multimodal data with feature level fusion. Furthermore, our approach achieved state-of-the-art performance on the urban CD problem posed by the OSCD dataset. Our Sentinel-1 SAR data and code are available onhttps://github.com/SebastianHafner/DS_UNet.

This research explores the potential of combining Sentinel-1 C-band Synthetic Aperture Radar (SAR) and Sentinel-2 MultiSpectral Instrument (MSI) data for Built-Up Area (BUA) mapping using deep learning. A lightweight U-Net model is trained using openly available building footprint reference data in North America and tested in four cities across three additional continents. The best test performance in terms of F1 score was achieved by the joint use of SAR and multispectral data (0.676), followed by multi-spectral (0.611) and SAR data (0.601). The developed fusion approach is particularly promising to distinguish BUA in low-density residential neighborhoods. Furthermore, our fusion approach compares favorably to the state-of-the-art in BUA mapping in the selected cities. However, associated with the diverse characteristics of human settlements around the world, considerable differences in accuracy among the test cities were observed. This indicates the need for more sophisticated fusion techniques to improve CNN model generalization and for adding more diverse training data.

In recent years, wildfires have become major devastating hazards that affect both public safety and the environment. Thus, agile detection of the wildfires is desirable to suppress wildfires in the early stage. Owing to the high temporal resolution, GOES-R satellites offer capabilities to obtain images every 15 minutes enabling a near real-time monitoring of wildfires. In this research, a time-series-based deep learning framework, composed of Gated Recurrent Units (GRU), is proposed to capture the emerging of the wildfire at early stage. By feeding the embedding of the coarse satellite imagery to Deep GRU network, the active fires are segmented out from the remote sensing imagery. The preliminary results show that proposed network can detect the wildfires earlier than the state-of-the-art fire product for 2020 wildfires in California and British Columbia, at the same time provide sufficiently high accuracy on the burned areas.

Synthetic Aperture Radar (SAR) satellite data are widely used to monitor deformation phenomena impacting the Earth's surface (e.g. landslides, glacier motions, subsidence, and volcano deformations) and infrastructures (e.g. bridges, dams, buildings). The analysis is generally performed using the Differential SAR Interferometry (DInSAR) technique that exploits the phase information of SAR data. However, this technique suffers for lack of coherence among the considered stack of images, and it can only be adopted to monitor slow deformation phenomena. In the field of geohazards monitoring and glacier melting, the Offset Tracking technique has been also widely investigated. This approach is based on the amplitude information only but it reaches worse accuracy compared to DInSAR. To overcome the limitations of DIn-SAR and Offset Tracking, in the last decade, a new technique called Imaging Geodesy has been investigated exploiting the amplitude information and the precise orbit of the modern SAR platforms (i.e. TerraSAR-X, COSMO-SkyMed). In this study, an investigation of using COSMO-SkyMed data for Earth surface monitoring was performed. The developed approach was applied to a set of imagery acquired over the Corvara (Northern Italy) area, which is affected by a fast landslide with yearly displacements up to meters. Specifically, two well identifiable and stable human-made Amplitude Persistent Scatterers (APSs) were considered to estimate the residual errors of COSMO-SkyMed sensor during the acquisition period between 2010 and 2015. Then, the same methodology was applied to estimate the displacement of a Corner Reflector (CR) located in the landslide area. Finally, the results were compared to the available GPS reference trend showing a good agreement.

The aim of this study is to compare and analyze the accuracy of freely available global DEMs. Four DEMs generated using optical and SAR satellite imagery - ASTER GDEM, SRTM DEM, ALOS AW3D30, Tandem-X 90m - were analyzed over the territories of four U.S. states (approximately 927000 km2) that are characterized by different morphologies and land covers. The accuracy assessment procedure was implemented within the Google Earth Engine (GEE) platform, designed to manage and analyze Geo Big Data. The outcomes highlight a good agreement among the statistical parameters at a global level for each wide area analyzed. The accuracy, as expected, decreases with the increase of the slopes, and ALOS AW3D30 displays the overall best performance, with an accuracy ranging between 2.5 m in flat areas and about 10 m in hilly/mountainous areas.

It is well known that SAR imagery is affected by SAR geometric distortions due to the SAR imaging process (i.e. Layover, Foreshortening and Shadows). Specially in mountainous areas these distortions affect large portions of the images and in some applications, these areas shouldn't be included in analysis. Using a foreshortening mask is a suitable solution, but finding the mask is challenging. The aim of this research is to exploit the fusion of Sentinel-1 multi-temporal images and SRTM DEM to produce a quasi-global foreshortening mask using the Google Earth Engine (GEE), cloud-based platform. The mean value of multi-temporal Sentinel-1 images is calculated. Then a local minimum algorithm finds probable foreshortening area. Aspect and slope information from SRTM DEM are used to refine Sentinel-1 derived foreshortening mask. The proposed method is tested in British Columbia (Canada), Everest Mountain (Nepal), and Mazandaran (Iran). The results demonstrate the reliability of proposed method to detect the foreshortening area.

The goal of this work is to exploit the slant-range measurements reaching centimetre accuracies using only the amplitude information of SAR data acquired by TerraSAR-X satellite sensor. The leading idea is to evaluate the positioning accuracy of well identifiable and stable natural and man-made Persistent Scatterers (PS's) along the SAR line of sight. New Earth observation SAR (Synthetic Aperture Radar) satellite sensors, as COSMO-SkyMed, TerraSAR-X and PAZ, acquire imagery on any point of the Earth with high resolutions, in terms of phase and amplitude value. Thanks to this higher amplitude resolution (up to 0.20 m pixel resolution in the Staring SpotLight mode for TerraSAR-X and PAZ) and to the use of on board dual frequency GPS receivers, which allows to determine the satellite orbit with an accuracy at few centimetres level, the SAR images offer the capability to achieve, in a global reference frame, positioning accuracies in the meter range and even better.

The present paper is focused on a potential method for unsupervised flood extent classification approach based, on the extraction of radar shadows from SAR simulated image, generated by means of image metadata and an available DSM covering the area of interest. The SAR simulation process is carried out using the SAR Simulator tool, implemented in the free and open SAR PlugIn developed for the Opticks platform. Different tests have been performed using SGF RADARSAT-2 imagery in order to assess the effectiveness of the proposed approach.

The actual high resolution optical and Synthetic Aperture Radar (SAR) satellite sensors offer interesting potentialities for Digital Surface Models (DSMs) generation. Both optical and SAR imagery are characterized by proper deformations and noise due to the different acquisition geometries and processes, which have to be duly taken into account during the DSM generation procedure in order to fully exploit the aforementioned potentialities. The aim of this work is to evaluate the performances of high resolution optical and SAR imagery for DSMs generation over the same testfield area where a dense network of GCPs and LiDAR DSM are available as ground truth data. The image processing and DSMs generation are carried out with the packages SISAR (Software Immagini Satellitari ad Alta Risoluzione) and SAT-PP (SATellite image Precision Processing) while an additional comparison is performed using PCI Geomatica 2012.

The availability of new high-resolution radar spaceborne sensors offers new interesting potentialities for the acquisition of data useful for the generation of Digital Surface Models (DSMs). Two different approaches may be used to generate DSMs from Synthetic Aperture Radar (SAR) data: the interferometric and the radargrammetric one. At present, the importance of the radargrammetric approach is rapidly growing due to the new high-resolution imagery [up to 1 m Ground Sample Distance (GSD)] which can be acquired by COSMO-SkyMed, TerraSAR-X and RADARSAT-2 in SpotLight mode. The defined and implemented model is related to COSMO- SkyMed SpotLight imagery in zero-Doppler geometry; it performs a 3-D orientation based on two range and two zero-Doppler equations, allowing for the least squares estimation of some calibration parameters, related to satellite position and velocity and to the range measure. The model has been implemented in SISAR (Software per Immagini Satellitari ad Alta Risoluzione), a scientific software developed at the Geodesy and Geomatic Institute of the University of Rome “La Sapienza”. Starting from this model, based on a geometric reconstruction, also a tool for the Rational Polynomial Coefficients (RPCs) generations has been implemented. To test the effectiveness of the new model, a stereo pair over the test sites of Merano (Northern Italy) has been orientated using the rigorous model and the RPCs one, and first results of radargrammetric DSM generation are presented; they display the possibility to reach an overall average accuracy of 3.5 m.

The availability of new high resolution optical and radar spaceborne sensors offers new interesting potentialities for the acquisition of data useful for the generation of Digital Surface Models (DSMs). The accuracy level of DSM is strictly related both to the image orientation and to the matching process. As regards the image orientation, remote sensing community usually adopts two different types of models for High Resolution Satellite Imagery (HRSI): the physical sensor models and the generalized sensor models also called rigorous and Rational Polynomial Functions (RPFs) models respectively. In a scientific software developed by the research group of Geodesy and Geomatic Area of the University of Rome "La Sapienza" both rigorous and RPFs models are implemented, with a specific tool for the terrain-independent Rational Polynomial Coefficients (RPCs) generation; the software manages the imagery acquired by several optical sensors (EROS A, Ikonos, QuickBird, Cartosat-1, WorldView-1, GeoEye-1) and by the Italian SAR constellation COSMO-SkyMed. In the same software a facility for image matching is embedded. The Area Base Matching (ABM) is used, combined with the orientation model re-parametrized in terms of RPCs. In the present work some examples of models application and DSM generation are analyzed and discussed.