Doctoral Meeting: 'Towards Resource-Efficient Deep Learning for Earth Observation: From Data Augmentation to Model Compression'

Earth Observation deep learning is constrained by two complementary efficiency challenges: the high cost and complexity of data acquisition, and the computational demands of large foundation models. This talk presents a unified perspective addressing both fronts. On the data side, we focus on high spatial resolution scenarios acquired by UAVs, where labeling is particularly expensive and datasets are typically small and imbalanced. To address this, we introduce a diffusion-based augmentation framework (EMViT-DDPM) tailored for multispectral imagery, which generates balanced, high-quality samples, removing bias while reducing dependence on costly data collection campaigns.

On the model side, we focus on reducing the size and computational load of Earth Observation foundation models through SIMPLER, a pre–fine-tuning method that identifies and removes redundant transformer layers via representation similarity analysis. This approach improves computational efficiency in both training and inference, significantly reducing model size while preserving most of the predictive performance. Crucially, this efficiency gain enhances the applicability of foundation models in resource-constrained environments, enabling their deployment on edge platforms such as satellites and UAVs.

Together, these contributions outline a pathway toward resource-efficient EO systems, where both the data pipeline and the model architecture are optimized to reduce costs without sacrificing accuracy.

• Supervisor: Dora Blanco Heras
• Moderator: Víctor Manuel Brea Sánchez