satellite remote sensing

Satellite imagery machine learning suffers from low data efficiency and poor out-of-distribution (Geo-OD) generalization across geographic regions. Method: We propose a multimodal modeling paradigm that integrates diverse geospatial data sources—including digital elevation models (DEM), land cover maps, and meteorological/environmental sensor data—into satellite image analysis. Through systematic comparison of hand-crafted fusion strategies (e.g., channel concatenation) versus learned fusion mechanisms (e.g., attention), we identify that simple, interpretable hand-crafted fusion yields superior robustness and performance under few-shot and cross-regional settings. Contribution/Results: Evaluated across multiple SatML benchmark tasks (classification, regression, segmentation), our approach significantly improves data utilization efficiency under label scarcity and enhances generalization to unseen geographic domains. It establishes a new paradigm for lightweight, reliable, and deployable remote sensing AI systems.

Interpreting spatiotemporal variability patterns across climate model ensembles remains challenging due to high dimensionality and structural complexity. Method: This paper proposes a visualization analytics workflow integrating Self-Organizing Maps (SOM) with Large Language Models (LLMs). SOM reduces dimensionality and clusters high-dimensional climate time series while preserving spatial structure; LLMs semantically interpret clustering outcomes, generating scientifically grounded, human-readable descriptions of climate patterns. The integrated system supports interactive exploration of variability magnitude, spatial configurations, and inter-model grouping relationships. Contribution/Results: Evaluated on precipitation projections over California and the U.S. Pacific Northwest, the method enables accurate identification of dominant variability modes, reveals model consensus and divergence, and produces expert-validated, interpretable insights. This work represents the first deep integration of LLMs into an SOM-driven climate ensemble analysis pipeline, significantly enhancing cognitive efficiency and interpretability of complex ensemble variability structures.

Soil moisture–precipitation coupling (SMPC) has long been challenging to quantify accurately due to strong nonlinearity in land–atmosphere interactions and weather-scale variability. To address this, we introduce high-dimensional model representation (HDMR) — a functional decomposition technique — to climate variable analysis for the first time, applied to the CONUS404 reanalysis dataset. This enables physically interpretable separation of direct effects, synergistic (second-order) interactions, and higher-order couplings. Results reveal that morning soil moisture over the U.S. Great Plains explains 40% of the variance in afternoon precipitation during summer; on rainy days, the first-order soil moisture effect enhances precipitation by up to 8 mm, while the second-order temperature–humidity interaction contributes an additional 3 mm. Our approach overcomes the persistent underestimation of land–atmosphere feedbacks inherent in conventional methods, substantially improving both the accuracy of SMPC quantification and the mechanistic interpretability of underlying processes.

This study addresses the limited scalability and reproducibility of existing crop yield prediction methods, which are often confined to specific crops or regions and lack standardized pipelines for multi-source environmental data. To overcome this, we propose UniCrop—the first general-purpose data engineering framework that decouples data specification from implementation—enabling automatic acquisition, cleaning, alignment, and fusion of over 200 environmental variables from sources such as Sentinel-1/2, MODIS, and ERA5-Land through simple configuration. The framework generates analysis-ready features for any location, crop, and time period. Coupled with mRMR feature selection and a LightGBM model, it achieves an R² of 0.6576 (RMSE = 465.1 kg/ha) using only 15 features across 557 rice fields; constrained ensemble learning further improves performance to R² = 0.6604, significantly enhancing model scalability, transparency, and practical utility.

To address the challenge of effectively assimilating sparse in-situ observations into full-atmosphere states for kilometer-scale weather forecasting initialization, this paper proposes a score-based generative data assimilation framework. First, an unconditional diffusion model is trained to learn the high-resolution atmospheric prior distribution—using HRRR analyses as ground truth. Then, sparse surface observations (e.g., precipitation, wind fields) are implicitly incorporated into the generative process via score matching. This work presents the first end-to-end, retraining-free generative assimilation method at the kilometer scale; it implicitly encodes multivariate physical constraints without explicit numerical physics modeling, thereby ensuring physical consistency and interpretability of generated fields. Experiments demonstrate that, given only 40 observation sites, the method reduces RMSE for surface variables at withheld sites by 10% relative to the HRRR baseline, while faithfully reproducing realistic meteorological structures—including fronts—with high spatial fidelity.

This study addresses the lack of systematic best practices in large-scale Earth observation (EO) mapping, which often introduces errors during data preprocessing, model training, inference deployment, and validation, thereby compromising the reliability and scientific credibility of map products. To remedy this, we propose the first end-to-end best practice framework for EO mapping, encompassing the entire workflow from satellite data acquisition to operational map delivery. The framework integrates six core components: EO data infrastructure, preprocessing, machine learning dataset construction, uncertainty quantification, map production and dissemination, and independent validation. Emphasizing the interdependence of these stages, it embeds uncertainty quantification and independent validation as integral elements. By synergizing machine learning, distributed computing, and geospatial validation techniques, the framework establishes a reproducible and scalable mapping pipeline that substantially enhances the quality, consistency, and scientific rigor of EO-derived maps, supported by open-source resources to foster community adoption.

Continuous monitoring of tropical cyclones is hindered by the long revisit intervals of microwave satellite imagery, sensor heterogeneity, irregular spatiotemporal sampling, and geographic misalignment across observations. To address these challenges, this work proposes a multi-source generative model that leverages self-supervised learning to perform spatiotemporal interpolation of heterogeneous microwave and infrared satellite images, enabling the first joint modeling of unaligned, non-uniformly sampled data. The approach incorporates a random source masking and reconstruction strategy to enhance generalization. Experiments demonstrate that the model significantly outperforms supervised baselines, as measured by reduced Continuous Ranked Probability Score (CRPS), with performance further improved through infrared data fusion. The generated outputs exhibit power spectra closely matching real observations, and their ensemble mean rivals that of deterministic forecast models.

This work proposes a machine learning–oriented paradigm for weather forecasting that reimagines the traditionally complex and closed operational systems to meet the demands of efficiency, openness, and collaboration in the era of artificial intelligence. By integrating agent-driven software engineering, open compressed data formats, shared validation workflows, interactive computing environments, and generative AI techniques, the framework systematically transforms model development, data utilization, computational management, and service delivery. Designed to equip meteorological and climate centers with future-ready infrastructure, it establishes robust data governance mechanisms, quality assurance protocols, and pathways for workforce skill transformation. The approach maintains scientific rigor while substantially enhancing the accessibility, efficiency, and interactivity of forecasting services.

This study addresses the limitations of traditional multi-source satellite precipitation estimation methods, which suffer from low computational efficiency, and the inflexibility of existing deep learning models in incorporating new sensor data. To overcome these challenges, the authors propose PRISMA, a novel and extensible generative framework for precipitation retrieval. PRISMA decouples modeling into an unconditional precipitation prior and independently trained sensor-conditional branches, enabling plug-and-play integration of new sensors without retraining the backbone network. By fusing infrared and microwave observations, the method achieves a 40.3% improvement in critical success index and a 22.6% reduction in root-mean-square error over microwave-covered regions. In typhoon case studies, absolute errors in storm core areas are reduced by up to 42.3%, with each inference requiring only approximately 37 seconds.

Existing foundation models for Earth observation struggle to effectively incorporate hyperspectral imagery (HSI), while specialized HSI models lack joint pretraining with multimodal remote sensing data. This work proposes a hierarchical Transformer architecture that, for the first time, enables unified pretraining of HSI alongside multispectral and SAR data through spectral tokenization, sensor-specific encoders, and a cross-sensor fusion module. The authors also introduce SpectralEarth-MM, a large-scale co-located multimodal dataset. Leveraging a JEPA-style joint embedding prediction objective, the model achieves state-of-the-art performance on both hyperspectral downstream tasks and general Earth observation benchmarks, significantly enhancing its generalization and multimodal fusion capabilities.