We study foundation models for Earth observation, focusing on transferable representations, multispectral imagery, cross-modal knowledge transfer, and efficient adaptation across sensors and tasks.

We develop supervised and unsupervised methods for detecting semantic changes in multi-temporal Earth-observation imagery, with an emphasis on robust design, generalization, and limited annotations.

We use time-series analysis of multispectral Sentinel-2 imagery to identify crops, with an emphasis on early classification and the temporal and spatial transferability of models.

We use deep learning and semantic segmentation to identify cultivated terraces in elevation data, addressing noisy labels, class imbalance, and transferability between regions.