This article presents Prithvi-EO-2.0, a new geospatial foundation model (GFM) 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 GFMs when benchmarked on remote sensing tasks from different domains and resolutions (i.e., from 0.1 to 15 m). The results demonstrate the versatility of the model in both classical Earth observation (EO) 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.

Remote sensing image super-resolution aims to recover fine spatial details lost during image acquisition, yet most existing deep learning methods often struggle to faithfully reconstruct high-frequency details and maintain spectral fidelity. To address these limitations, we build FC-GAN, a frequency-conditioned model that adapts frequency separation techniques to remote sensing data. FC-GAN employs a frequency-aware supervision approach that explicitly enforces consistency between the reconstructed outputs and the frequency characteristics of the original scene, ensuring that the synthesized textures and fine-scale spatial structures align with the ground-truth spatial power spectrum. In addition to spatial-domain metrics, we extend the standard frequency-domain evaluation framework from the meteorological downscaling literature to assess spectral fidelity using the Radially Averaged Power Spectral Density (RAPSD) and Log-Spectral Distance (RALSD). Experiments on three remote sensing datasets for ×4 upsampling demonstrate that FC-GAN improves reconstruction quality and better preserves frequency-domain statistics. The source code and model weights are publicly available at https://github.com/victor-nasc/FC-GAN.

TerraTorch is a fine-tuning and benchmarking toolkit for Geospatial Foundation Models built on PyTorch Lightning and tailored for satellite, weather, and climate data. It integrates domain-specific data modules, pre-defined tasks, and a modular model factory that pairs any backbone with diverse decoder heads. These components allow researchers and practitioners to fine-tune supported models in a no-code fashion by simply editing a training configuration. By consolidating best practices for model development and incorporating the automated hyperparameter optimization extension Iterate, TerraTorch reduces the expertise and time required to fine-tune or benchmark models on new Earth Observation use cases. Furthermore, TerraTorch directly integrates with GEO-Bench, allowing for systematic and reproducible benchmarking of Geospatial Foundation Models. TerraTorch is open sourced under Apache 2.0, available at https://github.com/IBM/terratorch, and can be installed via pip install terratorch.

The stochastic synthesis of extreme, rare climate scenarios is vital for risk and resilience models aware of climate change, directly impacting society in different sectors. However, creating high-quality variations of under-represented samples remains a challenge for several generative models. This paper investigates quantizing reconstruction losses for helping variational autoencoders (VAE) better synthesize extreme weather fields from conventional historical training sets. Building on the classical VAE formulation using reconstruction and latent space regularization losses, we propose various histogram-based penalties to the reconstruction loss that explicitly reinforces the model to synthesize under-represented values better. We evaluate our work using precipitation weather fields, where models usually strive to synthesize well extreme precipitation samples. We demonstrate that bringing histogram awareness to the reconstruction loss improves standard VAE performance substantially, especially for extreme weather events.

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

Significant progress in the development of highly adaptable and reusable Artificial Intelligence (AI) models is expected to have a significant impact on Earth science and remote sensing. Foundation models are pre-trained on large unlabeled datasets through self-supervision, and then fine-tuned for various downstream tasks with small labeled datasets. This paper introduces a first-of-a-kind framework for the efficient pre-training and fine-tuning of foundational models on extensive geospatial data. We have utilized this framework to create Prithvi, a transformer-based geospatial foundational model pre-trained on more than 1TB of multispectral satellite imagery from the Harmonized Landsat-Sentinel 2 (HLS) dataset. Our study demonstrates the efficacy of our framework in successfully fine-tuning Prithvi to a range of Earth observation tasks that have not been tackled by previous work on foundation models involving multi-temporal cloud gap imputation, flood mapping, wildfire scar segmentation, and multi-temporal crop segmentation. Our experiments show that the pre-trained model accelerates the fine-tuning process compared to leveraging randomly initialized weights. In addition, pre-trained Prithvi compares well against the state-of-the-art, e.g., outperforming a conditional GAN model in multi-temporal cloud imputation by up to 5pp (or 5.7%) in the structural similarity index. Finally, due to the limited availability of labeled data in the field of Earth observation, we gradually reduce the quantity of available labeled data for refining the model to evaluate data efficiency and demonstrate that data can be decreased significantly without affecting the model's accuracy. The pre-trained 100 million parameter model and corresponding fine-tuning workflows have been released publicly as open source contributions to the global Earth sciences community through Hugging Face.

Abstract Resampling‐based weather generators simulate new time series of weather variables by reordering the observed values such that the statistics of the simulated data are consistent with the observed ones. These generators are fully data‐driven, easy to implement, and capable of reproducing the dynamics among weather variables. However, although the simulated time series is new, the weather fields produced at arbitrary time steps are replicas of those found in observations, limiting the spatial variability of simulations and preventing the generation of extreme weather fields beyond the range of observed values. To address these limitations, we propose the integration of the Direct Sampling algorithm—a data‐driven method for producing simulations—into resampling‐based weather generators. By incorporating Direct Sampling as a post‐processing step on the outputs of the weather generator, we enhance the spatial variability of the generated weather fields and enable the generation of extreme weather fields. We introduce an approach for generating out‐of‐sample extreme weather fields using Direct Sampling. This method involves utilizing a set of control points in conjunction with Direct Sampling, where the values of these control points are informed by return period analysis. The proposed approach is validated using precipitation, temperature, and cloud cover weather fields in a region of northwest India. The experimental results confirm that Direct Sampling enhances the spatial variability of the weather fields and facilitates the generation of out‐of‐sample precipitation fields that accurately adhere to the spatial statistics provided by return precipitation level maps, as well as the observed precipitation weather field employed in the analysis.

A more accurate characterization of S2S extremes may result in great positive societal impact. Featurized S2S forecasts in the form of risk or extreme indices will aid in disaster response (especially for drought and flood events), inform disease outbreaks and heatwave onset, persistence, and decay. In this study, we identify a set of ECMWF-derived extreme weather indices that have spatio-temporal windows of opportunity for better-than-climatology skill. We report on the correlation between ECMWF-derived indices and ground-truth values. The selected indices can be calculated directly form probabilistic daily forecasts, or alternatively, by training specialized ML-models to process ensembles in a multi-task learning setup. Our goal is to find better approaches to communicate S2S climate risk by deploying a set of ECMWF-derived climate forecast products.

Recent years have seen increased interest in modeling future climate trends, especially from the point of view of accurately predicting, understanding and mitigating downstream impacts. For instance, current state-of-the-art process-based agriculture models rely on high-resolution climate data during the growing season for accurate estimation of crop yields. However, high-resolution climate data for future climates is unavailable and needs to be simulated, and that too for multiple possible climate scenarios, which becomes prohibitively expensive via traditional methods. Meanwhile, deep generative models leveraging the expressivity of neural networks have shown immense promise in modeling distributions in high dimensions. Here, we cast the problem of simulation of climate scenarios in a generative modeling framework. Specifically, we leverage the GAN (Generative Adversarial Network) framework for simulating synthetic climate scenarios. We condition the model by quantifying the degree of “extremeness" of the observed sample, which allows us to sample from different parts of the distribution. We demonstrate the efficacy of the proposed method on the CHIRPS precipitation dataset.

One of the consequences of climate change is anobserved increase in the frequency of extreme cli-mate events. That poses a challenge for weatherforecast and generation algorithms, which learnfrom historical data but should embed an often un-certain bias to create correct scenarios. This paperinvestigates how mapping climate data to a knowndistribution using variational autoencoders mighthelp explore such biases and control the synthesisof weather fields towards more extreme climatescenarios. We experimented using a monsoon-affected precipitation dataset from southwest In-dia, which should give a roughly stable pattern ofrainy days and ease our investigation. We reportcompelling results showing that mapping complexweather data to a known distribution implementsan efficient control for weather field synthesis to-wards more (or less) extreme scenarios.

Climate change is making extreme weather more extreme. Given the inherent uncertainty of long-term climate projections, there is growing need for rapid, plausible “what-if” climate scenarios to help users understand climate exposure and examine resilience and mitigation strategies. Since the 1980s, such “what-if” scenarios have been created using stochastic weather generators. However, it is very challenging for traditional weather generation algorithms to create realistic extreme climate scenarios because the weather data being modeled is highly imbalanced, contains spatiotemporal dependencies and has extreme weather events exacerbated by a changing climate. There are few works comparing and evaluating stochastic multisite (i.e., gridded) weather generators, and no existing work that compares promising deep learning approaches for weather generation with classical stochastic weather generators. We will present the culmination of a multi-year effort to perform a systematic evaluation of stochastic weather generators and deep generative models for multisite precipitation synthesis. Among other things, we show that variational auto-encoders (VAE) offer an encouraging pathway for efficient and controllable climate scenario synthesis – especially for extreme events. Our proposed VAE schema selects events with different characteristics in the normalized latent space (from rare to common) and generates high-quality scenarios using the trained decoder. Improvements are provided via latent space clustering and bringing histogram-awareness to the VAE loss. This research will serve as a guide for improving the design of deep learning architectures and algorithms for application in Earth science, including feature representation and uncertainty quantification of Earth system data and the characterization of so-called “grey swan” events.

Wildfire events have caused severe losses in many places around the world and are expected to increase with climate change. Throughout the years many technologies have been developed to identify fire events early on and to simulate fire behavior once they have started. Another particularly helpful technology is fire risk indices, which use weather forcing to make advanced predictions of the risk of fire. Predictions of fire risk indices can be used, for instance, to allocate resources in places with high risk. These indices have been developed over the years as empirical models with parameters that were estimated in lab experiments and field tests. These parameters, however, may not fit well all places where these models are used. In this paper we propose a novel implementation of one index (NFDRS IC) as a differentiable function in which one can optimize its internal parameters via gradient descent. We leverage existing machine learning frameworks (PyTorch) to construct our model. This approach has two benefits: (1) the NFDRS IC parameters can be improved for each region using actual observed fire events, and (2) the internal variables remain intact for interpretations by specialists instead of meaningless hidden layers as in traditional neural networks. In this paper we evaluate our strategy with actual fire events for locations in the USA and Europe.

Climate change is making extreme weather more extreme. Given the inherent uncertainty of long-term climate projections, there is growing need for rapid, plausible “what-if” climate scenarios to help users understand climate exposure and examine resilience and mitigation strategies. Since the 1980s, such “what-if” scenarios have been created using stochastic weather generators. However, it is very challenging for traditional weather generation algorithms to create realistic extreme climate scenarios because the weather data being modeled is highly imbalanced, contains spatiotemporal dependencies and has extreme weather events exacerbated by a changing climate. There are few works comparing and evaluating stochastic multisite (i.e., gridded) weather generators, and no existing work that compares promising deep learning approaches for weather generation with classical stochastic weather generators. We will present the culmination of a multi-year effort to perform a systematic evaluation of stochastic weather generators and deep generative models for multisite precipitation synthesis. Among other things, we show that variational auto-encoders (VAE) offer an encouraging pathway for efficient and controllable climate scenario synthesis – especially for extreme events. Our proposed VAE schema selects events with different characteristics in the normalized latent space (from rare to common) and generates high-quality scenarios using the trained decoder. Improvements are provided via latent space clustering and bringing histogram-awareness to the VAE loss. This research will serve as a guide for improving the design of deep learning architectures and algorithms for application in Earth science, including feature representation and uncertainty quantification of Earth system data and the characterization of so-called “grey swan” events.

<p>The need to build reliable weather forecasting systems for subseasonal to seasonal (S2S) timescales has never been greater as the world continues to experience increased numbers of extreme weather events. This study addresses the skill gap between numerical weather prediction (NWP) and seasonal forecasting by proposing a daily probabilistic forecast model that predicts 2-meter temperature and total precipitation on a global scale. It combines multimodal data (e.g., physics-based ensembles, climate modes, recent climatology) into feature vectors given as inputs to three ML models: Extreme Gradient Boosting, U-Net, and Natural Gradient Boosting. We use Bayesian hyperparameter search and leave-one-year-out RPSS cross-validated scores to accelerate learning and ensure generalizability. Our method consistently outperforms both ECMWF 46-day forecasts and climatology. We find that augmenting physics-based issued forecasts with other sources of predictability greatly improves the performance of the underlying dynamical models. We hope that by improving the physics-based probabilistic forecasts, we will unlock skill in predicting climate extremes-oriented indices. Subsequent representation learning models may be trained to efficiently navigate the ensembles' uncertainty space and estimate the likelihood of extreme events.</p>

Future climate change scenarios are usually hypothesized using simulations from weather generators. However, there only a few works comparing and evaluating promising deep learning models for weather generation against classical approaches. This study shows preliminary results making such evaluations for the multisite precipitation synthesis task. We compared two open-source weather generators: IBMWeathergen (an extension of the Weathergen library) and RGeneratePrec, and two deep generative models: GAN and VAE, on a variety of metrics. Our preliminary results can serve as a guide for improving the design of deep learning architectures and algorithms for the multisite precipitation synthesis task.

The construction of conceptual geological models is an essential task in petroleum exploration, especially during the early stages of investment, when evidence about the subsurface is limited. In this task, geoscientists recreate the most likely geological scenarios that led to potential accumulation of reserves in a target block, based on past experience, historical analogues, and interpreted "signatures" that were left in the data by physical processes. Due to cognitive constraints, this task has traditionally focused on the single most likely conceptual scenario, or at most, a reduced set of scenarios chosen a priori via ad-hoc methods, which often lead to improper block valuation and severe money losses. In this work, we propose a probabilistic framework for reasoning about conceptual geological scenarios that helps domain experts maintain multiple hypotheses throughout the exploration program. The framework is extensible and can be instantiated automatically from simple knowledge templates, a form of "knowledge standard" in the company. We show how the acquired knowledge can be leveraged for uncertainty mitigation using concepts from information theory, and assess the framework qualitatively in a real case study.

An impact of climate change is the increase in frequency and intensity of extreme precipitation events. However, confidently predicting the likelihood of extreme precipitation at seasonal scales remains an outstanding challenge. Here, we present an approach to forecasting the quantiles of the maximum daily precipitation in each week up to six months ahead using the temporal fusion transformer (TFT) model. Through experiments in two regions, we compare TFT predictions with those of two baselines: climatology and a calibrated ECMWF SEAS5 ensemble forecast (S5). Our results show that, in terms of quantile risk at six month lead time, the TFT predictions significantly outperform those from S5 and show an overall small improvement compared to climatology. The TFT also responds positively to departures from normal that climatology cannot.

In many important application domains, such as text categorization, biomolecular analysis, scene or video classification and medical diagnosis, instances are naturally associated with more than one class label, giving rise to multi-label classification problems. This has led, in recent years, to a substantial amount of research in multi-label classification. More specifically, feature selection methods have been developed to allow the identification of relevant and informative features for multi-label classification. This work presents a new feature selection method based on the lazy feature selection paradigm and specific for the multi-label context. Experimental results show that the proposed technique is competitive when compared to multi-label feature selection techniques currently used in the literature, and is clearly more scalable, in a scenario where there is an increasing amount of data.

Extreme weather events have an enormous impact on society and are expected to become more frequent and severe with climate change. In this context, resilience planning becomes crucial for risk mitigation and coping with these extreme events. Machine learning techniques can play a critical role in resilience planning through the generation of realistic extreme weather event scenarios that can be used to evaluate possible mitigation actions. This paper proposes a modular framework that relies on interchangeable components to produce extreme weather event scenarios. We discuss possible alternatives for each of the components and show initial results comparing two approaches on the task of generating precipitation scenarios.

Operational forecasting centers are investing in decadal (1-10 year) forecast systems to support long-term decision making for a more climate-resilient society. One method that has previously been employed is the Dynamic Mode Decomposition (DMD) algorithm - also known as the Linear Inverse Model - which fits linear dynamical models to data. While the DMD usually approximates non-linear terms in the true dynamics as a linear system with random noise, we investigate an extension to the DMD that explicitly represents the non-linear terms as a neural network. Our weight initialization allows the network to produce sensible results before training and then improve the prediction after training as data becomes available. In this short paper, we evaluate the proposed architecture for simulating global sea surface temperatures and compare the results with the standard DMD and seasonal forecasts produced by the state-of-the-art dynamical model, CFSv2.