Geospatial Analysis and Foundation Models

The field of geospatial analysis is moving towards more efficient and robust foundation model adaptation, with a focus on innovative decoder architectures and pre-training frameworks. Recent developments have introduced dynamic adaptive regularization networks, contrastive learning with dynamic instances and contour consistency, and training-free open-vocabulary segmentation frameworks. These advancements have shown exceptional performance across various geospatial tasks, including land-cover classification, roadway crash risk assessment, and underwater segmentation. Noteworthy papers include DARN, which achieves state-of-the-art performance on the GeoBench benchmark, and DI3CL, which develops a general-purpose foundation model for SAR land-cover classification. Additionally, Earth2Ocean and NERVE have demonstrated significant performance improvements in underwater segmentation and open-vocabulary semantic segmentation, respectively.

Sources

DARN: Dynamic Adaptive Regularization Networks for Efficient and Robust Foundation Model Adaptation

Beta Distribution Learning for Reliable Roadway Crash Risk Assessment

DI3CL: Contrastive Learning With Dynamic Instances and Contour Consistency for SAR Land-Cover Classification Foundation Model

Exploring the Underwater World Segmentation without Extra Training

LandSegmenter: Towards a Flexible Foundation Model for Land Use and Land Cover Mapping

NERVE: Neighbourhood & Entropy-guided Random-walk for training free open-Vocabulary sEgmentation

Empowering DINO Representations for Underwater Instance Segmentation via Aligner and Prompter

Built with on top of