This paper presents Prithvi-EO-2.0, a new geospatial foundation model that offers significant improvements over its predecessor, Prithvi-EO-1.0. Trained on 4.2 million global time series samples from NASA’s Harmonized Landsat and Sentinel-2 data archive at 30-m resolution, the new model incorporates temporal and location embeddings for enhanced performance across various geospatial tasks. Through extensive benchmarking with GEO-Bench, the model outperforms the previous Prithvi-EO model by 8% across a range of tasks. It also outperforms six other geospatial foundation models when benchmarked on remote sensing tasks from different domains and resolutions (i.e. from 0.1 m to 15 m). The results demonstrate the versatility of the model in both classical Earth observation and high-resolution applications. Early involvement of end-users and subject matter experts (SMEs) allowed constant feedback on model and dataset design, enabling customization across diverse SME-led applications in disaster response, land cover and crop mapping, and ecosystem dynamics monitoring. Prithvi-EO-2.0 is available as an open-source model on Hugging Face and IBM TerraTorch, with additional resources on GitHub. The project exemplifies the Trusted Open Science approach embraced by all involved organizations.

High-Performance Computing (HPC) has recently been attracting more attention in remote sensing applications due to the challenges posed by the increased amount of open data that are produced daily by Earth Observation (EO) programs. The unique parallel computing environments and programming techniques that are integrated in HPC systems are able to solve large-scale problems such as the training of classification algorithms with large amounts of Remote Sensing (RS) data. This paper shows that the training of state-of-the-art deep Convolutional Neural Networks (CNNs) can be efficiently performed in distributed fashion using parallel implementation techniques on HPC machines containing a large number of Graphics Processing Units (GPUs). The experimental results confirm that distributed training can drastically reduce the amount of time needed to perform full training, resulting in near linear scaling without loss of test accuracy.

This paper presents Sen4Map, a large-scale benchmark dataset designed to enhance the capability of generating land-cover maps using Sentinel-2 data. Comprising non-overlapping $64\times 64$ patches extracted from Sentinel-2 time series images, the dataset spans 335,125 geo-tagged locations across the European Union. These locations are associated with detailed land-cover and land-use information gathered by expert surveyors in 2018. Unlike most existing large datasets available in the literature, the presented database provides: (1) a detailed description of the land-cover and land-use properties of each sampled area; (2) independence of scale, as it is associated with reference data collected in-situ by expert surveyors; (3) the ability to test both temporal and spatial classification approaches because of the availability of time series of $64\times 64$ patches associated with each labeled sample; and (4) samples were collected following a stratified random sample design to obtain a statistically representative spatial distribution of land-cover classes throughout the European Union. To showcase the properties and challenges offered by Sen4Map, we benchmarked the current state-of-the-art land-cover classification approaches. The dataset and code can be downloaded at: https://datapub.fz-juelich.de/sen4map.

Over the past decades, there has been an explosion in the amount of available Earth observation (EO) data. The unprecedented coverage of Earth’s surface and atmosphere by satellite imagery has resulted in large volumes of data that must be transmitted to ground stations, stored in data centers, and distributed to end users. Modern Earth system models (ESMs) face similar challenges, operating at high spatial and temporal resolutions, producing petabytes of data per simulated day. Data compression has gained relevance over the past decade, with neural compression (NC) emerging from deep learning and information theory, making EO data and ESM outputs ideal candidates because of their abundance of unlabeled data. In this review, we outline recent developments in NC applied to geospatial data. We introduce the fundamental concepts of NC, including seminal works in its traditional applications to image and video compression domains with a focus on lossy compression. We discuss the unique characteristics of EO and ESM data, contrasting them with “natural images,” and we explain the additional challenges and opportunities they present. Additionally, we review current applications of NC across various EO modalities and explore the limited efforts in ESM compression to date. The advent of self-supervised learning (SSL) and foundation models (FMs) has advanced methods to efficiently distill representations from vast amounts of unlabeled data. We connect these developments to NC for EO, highlighting the similarities between the two fields and elaborate on the potential of transferring compressed feature representations for machine-to-machine communication. Based on insights drawn from this review, we devise future directions relevant to applications in EO and ESMs.

The combination of data acquired by Landsat-8 and Sentinel-2 Earth Observation (EO) missions produces dense Time Series (TSs) of multispectral images that are essential for monitoring the dynamics of land-cover and land-use classes across the Earths surface with high temporal resolution. However, the optical sensors of the two missions have different spectral and spatial properties, thus they require a harmonization processing step before they can be exploited in Remote Sensing (RS) applications. In this work, we propose a workflow based on a Deep Learning (DL) approach to harmonize these two products developed and deployed on an High-Performance Computing (HPC) environment. In particular, we use a multispectral Generative Adversarial Network (GAN) with a U-Net generator and a PatchGan discriminator to integrate existing Landsat-8 TSs with data sensed by the Sentinel-2 mission. We show a qualitative and quantitative comparison with an existing physical method (NASA Harmonized Landsat and Sentinel (HLS)) and analyze original and generated data in different experimental setups with the support of spectral distortion metrics. To demonstrate the effectiveness of the proposed approach, a crop type mapping task is addressed using the harmonized dense TS of images, which achieved an Overall Accuracy (OA) of 87.83% compared to 81.66% of the state-of-the-art method.

Large-scale crop yield mapping has long relied on the Moderate Resolution Imaging Spectrometer (MODIS) due to its high temporal resolution and consistent atmospheric correction. The Sentinel-2 constellation, with its finer spatial resolution and vegetation-sensitive spectral bands, now offers new opportunities for regional- and field-scale yield prediction—especially as MODIS nears the end of its operational life. However, it remains unclear whether Sentinel-2 can ensure continuity of MODIS-based estimates across diverse agricultural regions. We present a regional sensor-to-sensor comparison of MODIS and Sentinel-2 for crop yield prediction using matched spatiotemporal inputs across two agro-ecological zones. Using reproducible regression workflows, we demonstrate that Sentinel-2 captures finer spatial variation in crop phenology and consistently outperforms MODIS in terms of predictive accuracy. For cotton, Sentinel-2 achieved an RMSE of 123.52 lb/acre and R² of 0.76, versus MODIS with 129.20 lb/acre and R² of 0.74. For corn, Sentinel-2 achieved 8.40 Bu/acre and 0.79, outperforming MODIS at 8.69 and 0.66, respectively. SHAP analysis identifies Enhanced Vegetation Index (EVI), Fraction of Photosynthetically Active Radiation (FPAR), and Leaf Area Index (LAI) as key predictors across both sensors. Despite its lower temporal frequency, Sentinel-2 delivers robust, regionally consistent estimates, supporting its suitability as a successor to MODIS for operational crop monitoring.