Researchers introduced SomBench, a benchmark dataset aimed at machine learning applications in lunar science. The work addresses the difficulty of combining observations from multiple lunar orbital missions, which differ in sampling, projection, and instrument characteristics. It is intended to give researchers a common basis for evaluating models on lunar data.
Researchers introduce a multimodal, multiresolution foundation model trained from scratch for lunar remote sensing. It was pretrained on SomBench, a geographically partitioned dataset of roughly two million co-registered tile bundles covering 11 sensor modalities at two spatial resolutions of 1 m/pixel. The work targets general-purpose representation learning for planetary surface analysis.
Researchers examined whether simulated snowpack conditions from the SNOWPACK model can serve as a data-driven predictor of avalanche activity detected by satellite imagery. The work targets regions where direct field observations are too sparse to support conventional forecasting. It points toward filling observational gaps in mountain avalanche monitoring with model-based estimates.
A new arXiv paper applies hyperbolic geometry to open-world object detection in satellite and aerial imagery. The work targets the fact that remote-sensing object categories carry hidden hierarchical structure, which standard Euclidean embedding spaces struggle to represent. The method is designed to flag unknown objects and incrementally absorb them into the model once annotations become available.