High-quality pixel-level annotations are essential for the semantic segmentation of remote sensing imagery. However, such labels are expensive to obtain and are often affected by noise due to the labor-intensive and time-consuming nature of pixel-wise annotation, which makes it challenging for human annotators to label every pixel accurately. Annotation errors can significantly degrade the performance and robustness of modern segmentation models, motivating the need for reliable mechanisms to identify and quantify noisy training samples. This paper introduces a novel data-centric benchmark, together with a new, publicly available binary building segmentation dataset. This specific task serves as a representative and practically relevant testbed that enables controlled experimentation with different annotation perturbations. Furthermore, we also introduce two techniques for identifying, quantifying, and ranking training samples according to their level of label noise in remote sensing semantic segmentation. The proposed methods leverage complementary strategies based on model uncertainty, prediction consistency, and representation analysis, and consistently outperform established baselines across a range of experimental settings. The outcomes of this work are publicly available at https://github.com/keillernogueira/label_noise_segmentation.

Earth Observation (EO) systems combined with Artificial Intelligence (AI) techniques have significantly advanced in recent years. The emergence and success of foundational models (FMs), such as ChatGPT and DALL-E 2, based on Large Language Models (LLMs), have influenced the development of the so-called "Geo'' Foundational Models (GFMs). These models are trained using Self-Supervised Learning (SSL) strategy on broad datasets and can be adapted to a wide variety of downstream tasks, such as flood mapping, Building Damage Classification (BDC), and Land Cover (LC) classification. This versatility addresses complex issues in Earth science, offering a significant advantage compared to traditional Deep Learning (DL) or Machine Learning (ML) models, which are typically restricted to one or several data types and optimized for a single task, such as classification or change detection. With the launch of GFM Clay, which was initially pre-trained for segmentation, classification, and biomass information extraction using a variety of sensors such as Sentinel-1, Sentinel-2, and Landsat, we fine-tune the model for the task of building damage assessment (BDA) in Beira, Mozambique. Therefore, in this study, the model is fine-tuned to the EDDA dataset, a high-resolution drone dataset released in October 2023, which captures building damage levels after Tropical Cyclone (TC) Idai devastated nearly 90% of Beira, the fourth-largest city in Mozambique. The fine-tuned Clay model achieved a validation mean intersection over union (mIoU) score of 0.829 and an F1 score of 0.869, demonstrating strong performance even with imbalanced data, compared to a fully supervised U-Net model, that we trained using the same dataset, and attained an mIoU of 0.567 and F1-score of 0.717. The qualitative results observed are highly acceptable, underscoring the transferability of the Clay Model to different tasks, datasets, and regions. These capabilities are crucial assets for its potential globalization.

Wildfire burned-area mapping is essential for damage assessment, emissions modeling, and understanding fire–climate interactions across diverse ecological regions. Recent geospatial foundation models provide strong general-purpose representations for satellite imagery, yet there is still no clear understanding of how to efficiently adapt these models for downstream Earth observation tasks, particularly under geographic and temporal domain shift. This study evaluates three state-of-the-art Geospatial Foundation Models (GFMs) - Terramind, DINOv3, and Prithvi-v2 - for burned-area mapping across the United States and Canada using Sentinel-2 data. Leveraging 3,820 wildfire events from 2017–2023, we conduct spatial and temporal generalization tests across diverse biomes. We systematically compare full fine-tuning, decoder-only fine-tuning, and Low-Rank Adaptation (LoRA) for adapting each model. Across all experiments, LoRA provides the strongest cross-domain generalization while updating less than 1% of parameters, demonstrating a favorable trade-off between accuracy and efficiency. Prithvi-v2 with LoRA achieves the highest overall accuracy and the larger improvement compare to full fine-tuning. These findings indicate that geospatial foundation models, when adapted using lightweight parameter-efficient methods such as LoRA, offer a robust and scalable solution for large-scale burned-area mapping. Code is available at https://github.com/alishibli97/wildfire-lora-gfm.

Geospatial foundation models (GFMs) have emerged as powerful tools for extracting representations from Earth observation (EO) data, but their evaluation remains inconsistent and narrow. Existing works often evaluate on suboptimal downstream datasets and tasks that are often too easy or too narrow, limiting the usefulness of the evaluations to assess the real-world applicability of GFMs. Additionally, there is a distinct lack of diversity in current evaluation protocols, which fail to account for the multiplicity of image resolutions, sensor types, and temporalities, which further complicates the assessment of GFM performance. In particular, most existing benchmarks are geographically biased toward North America and Europe, questioning the global applicability of GFMs. To overcome these challenges, we introduce PANGAEA, a standardized evaluation protocol that covers a diverse set of datasets, tasks, resolutions, sensor modalities, and temporalities. It establishes a robust and widely applicable benchmark for GFMs. We evaluate the most popular GFMs openly available on this benchmark and analyze their performance across several domains. In particular, we compare these models to supervised baselines [e.g., UNet and vanilla vision-Transformer (ViT)] and assess their effectiveness when faced with limited labeled data. Our findings highlight the limitations of GFMs under different scenarios, showing that they do not consistently outperform supervised models. PANGAEA is designed to be highly extensible, allowing for the seamless inclusion of new datasets, models, and tasks in future research. By releasing the evaluation code and benchmark, we aim to enable other researchers to replicate our experiments and build upon our work, fostering a more principled evaluation protocol for large pretrained geospatial models. The code is available at https://github.com/VMarsocci/pangaea-bench.

Building heights are crucial for sustainable urban planning and monitoring. While traditional methods use airborne stereo images and LiDAR data for accurate height estimation, their large-scale application is costly and slow, limiting the ability to conduct frequent large-scale monitoring. In contrast, satellite data offers a scalable alternative, further improved with Deep Learning (DL) models. However, the lack of representative open-source training datasets has constrained the progress in this field. In this paper, we introduce M4Heights, a multi-modal, multi-resolution, and multi-temporal dataset designed for building height estimation, spanning diverse architectural styles, urban densities, and terrain complexities across Estonia, Netherlands, and Switzerland. The dataset includes approximate to 1 million images, comprising time series of Sentinel-1 SAR and Sentinel-2 MSI satellite data, high-resolution aerial orthophotos, and high-quality building height reference maps. Additionally, M4Heights provides the largest associated multi-image super-resolution dataset to enhance height estimation accuracy. Our dataset supports a range of modeling approaches, offers extensibility to new geographic regions and provides opportunities to advance the development of DL models for building height estimation.

Since traditional inspection methods of structures are time-consuming and prone to human errors, many researchers have investigated the possibility of using various deep learning models to automate damage detection and, in particular, crack detection. However, deep learning models require a large amount of training data to predict reasonably accurate results. Creating a dataset with segmented cracks is time-consuming, and the aim of this paper is, therefore, to present a semi-automated labelling process of cracks. This has the potential to greatly decrease the time spent creating datasets.

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.

Tunnels are essential for infrastructure and require regular visual inspections by trained operators to identify defects like cracks and water ingress. This process is time-consuming, prone to human error, and requires tunnel closures. The need for efficient inspection solutions has increased as tunnel networks expand and age. Recent advancements in mobile mapping systems equipped with geomatic sensors, such as cameras and LiDAR sensors, have significantly enhanced data collection. These systems allow for rapid data acquisition, reducing tunnel downtime and enabling the generation of digital twins for remote inspections, which improves knowledge transfer. However, damage detection remains manual even if deep learning methods have been extensively investigated to automate defect detection using high-resolution images collected with mobile mapping systems. This paper is part of the TACK-II project and focuses on mobile mapping systems, providing a detailed analysis of the parameters that influence their design for effective data collection in tunnels.

Accurate building height estimation is essential to support urbanization monitoring, environmental impact analysis and sustainable urban planning. However, conducting large-scale building height estimation remains a significant challenge. While deep learning (DL) has proven effective for large-scale mapping tasks, there is a lack of advanced DL models specifically tailored for height estimation, particularly when using open-source Earth observation data. In this study, we propose T-SwinUNet, an advanced DL model for large-scale building height estimation leveraging Sentinel-1 SAR and Sentinel-2 multispectral time series. T-SwinUNet model contains a feature extractor with local/global feature comprehension capabilities, a temporal attention module to learn the correlation between constant and variable features of building objects over time and an efficient multitask decoder to predict building height at 10 m spatial resolution. The model is trained and evaluated on data from the Netherlands, Switzerland, Estonia, and Germany, and its generalizability is evaluated on an out-of-distribution (OOD) test set from ten additional cities from other European countries. Our study incorporates extensive model evaluations, ablation experiments, and comparisons with established models. T-SwinUNet predicts building height with a Root Mean Square Error (RMSE) of 1.89 m, outperforming state-of-the-art models at 10 m spatial resolution. Its strong generalization to the OOD test set (RMSE of 3.2 m) underscores its potential for low-cost building height estimation across Europe, with future scalability to other regions. Furthermore, the assessment at 100 m resolution reveals that T-SwinUNet (0.29 m RMSE, 0.75 R2) also outperformed the global building height product GHSL-Built-H R2023A product(0.56 m RMSE and 0.37 R2). Our implementation is available at: https://github.com/RituYadav92/Building-Height-Estimation.

This study evaluates the effectiveness of integrating high-resolution remote sensing (RS) data with machine learning (ML) techniques to identify criminogenic environments in urban areas. We employ an unsupervised ISO clustering method to classify land cover from aerial imagery, thereby capturing fine-scale environmental details that are often overlooked in traditional analyses. These clusters are linked to both crime and non-crime events through a presence/absence (case–control) framework, a methodology adapted from species distribution studies, which enables a micro-environmental examination of crime locations. In addition to RS-derived land-cover predictors, the study incorporates socio-economic and demographic variables, as well as a centrality indicator that proxies the intensity of urban activity. A Random Forest classifier is utilized to model the likelihood of street theft incidents based on these predictors. The model achieves robust performance, with an F1-score of 0.88 ± 0.03 as determined by K-fold cross-validation. To enhance model interpretability, SHapley Additive exPlanations (SHAP) is applied. The findings of this research demonstrate that integrating RS data with ML techniques offers a valuable tool for identifying and mapping criminogenic environments. The resulting risk map of Stockholm highlights key urban areas with elevated street theft risk, offering guidance for targeted crime prevention and urban planning strategies. While our workflow simplifies some technical steps compared to other RS + ML pipelines, it still requires GIS and ML competence to implement effectively. This approach reduces, but does not eliminate, sensitivity to spatial unit choice (MAUP) and spatial data dependencies.