Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rapid Adaptation of Earth Observation Foundation Models for Segmentation

Karthick Panner Selvam, Raul Ramos-Pollan, Freddie Kalaitzis

TL;DR

This work tackles the problem of rapidly adapting large Earth Observation transformer models for flood segmentation under limited computational resources. It applies Low-Rank Adaptation (LoRA) to the Clay EO-FM encoder, inserting low-rank updates in attention modules while using a custom decoder, and optimizes a joint BCE-Dice loss to produce precise flood masks from Sentinel-1 SAR data. The findings show that LoRA with a mid-range rank (r=256) yields the best in-distribution performance, outperforming both frozen-encoder baselines and infeasible full fine-tuning, with substantial parameter efficiency. OOD evaluation on the Luxembourg flood event reveals generalization improvements with higher LoRA ranks, though challenges remain in transferring to unseen geographical contexts, underscoring the need for robust, resource-aware deployment in disaster response scenarios.

Abstract

This study investigates the efficacy of Low-Rank Adaptation (LoRA) in fine-tuning Earth Observation (EO) foundation models for flood segmentation. We hypothesize that LoRA, a parameter-efficient technique, can significantly accelerate the adaptation of large-scale EO models to this critical task while maintaining high performance. We apply LoRA to fine-tune a state-of-the-art EO foundation model pre-trained on diverse satellite imagery, using a curated dataset of flood events. Our results demonstrate that LoRA-based fine-tuning (r-256) improves F1 score by 6.66 points and IoU by 0.11 compared to a frozen encoder baseline, while significantly reducing computational costs. Notably, LoRA outperforms full fine-tuning, which proves computationally infeasible on our hardware. We further assess generalization through out-of-distribution (OOD) testing on a geographically distinct flood event. While LoRA configurations show improved OOD performance over the baseline. This work contributes to research on efficient adaptation of foundation models for specialized EO tasks, with implications for rapid response systems in disaster management. Our findings demonstrate LoRA's potential for enabling faster deployment of accurate flood segmentation models in resource-constrained, time-critical scenarios.

Rapid Adaptation of Earth Observation Foundation Models for Segmentation

TL;DR

This work tackles the problem of rapidly adapting large Earth Observation transformer models for flood segmentation under limited computational resources. It applies Low-Rank Adaptation (LoRA) to the Clay EO-FM encoder, inserting low-rank updates in attention modules while using a custom decoder, and optimizes a joint BCE-Dice loss to produce precise flood masks from Sentinel-1 SAR data. The findings show that LoRA with a mid-range rank (r=256) yields the best in-distribution performance, outperforming both frozen-encoder baselines and infeasible full fine-tuning, with substantial parameter efficiency. OOD evaluation on the Luxembourg flood event reveals generalization improvements with higher LoRA ranks, though challenges remain in transferring to unseen geographical contexts, underscoring the need for robust, resource-aware deployment in disaster response scenarios.

Abstract

This study investigates the efficacy of Low-Rank Adaptation (LoRA) in fine-tuning Earth Observation (EO) foundation models for flood segmentation. We hypothesize that LoRA, a parameter-efficient technique, can significantly accelerate the adaptation of large-scale EO models to this critical task while maintaining high performance. We apply LoRA to fine-tune a state-of-the-art EO foundation model pre-trained on diverse satellite imagery, using a curated dataset of flood events. Our results demonstrate that LoRA-based fine-tuning (r-256) improves F1 score by 6.66 points and IoU by 0.11 compared to a frozen encoder baseline, while significantly reducing computational costs. Notably, LoRA outperforms full fine-tuning, which proves computationally infeasible on our hardware. We further assess generalization through out-of-distribution (OOD) testing on a geographically distinct flood event. While LoRA configurations show improved OOD performance over the baseline. This work contributes to research on efficient adaptation of foundation models for specialized EO tasks, with implications for rapid response systems in disaster management. Our findings demonstrate LoRA's potential for enabling faster deployment of accurate flood segmentation models in resource-constrained, time-critical scenarios.
Paper Structure (11 sections, 13 equations, 3 figures, 2 tables)

This paper contains 11 sections, 13 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Earth Observation Foundation Models
  4. Theoretical Analysis of Low-Rank Adaptation (LoRA)
  5. Methodology
  6. Problem Formulation
  7. Model Architecture
  8. Training Objective
  9. Experiments and Results
  10. Out-of-Distribution Generalization
  11. Conclusion

Figures (3)

  • Figure 1: Dataset composition for flood detection. Each sample includes pre-event and post-event Sentinel-1 SAR imagery (VV and VH polarizations for both ascending and descending orbits) along with corresponding flood extent labels. Pre: pre-event; Post: post-event; VV: vertical transmit-vertical receive polarization; VH: vertical transmit-horizontal receive polarization.
  • Figure 2: Visualization of patch embeddings from the pre-trained Clay EO-FM encoder using t-SNE projection. The projections consistently demonstrate clear clustering of water and land regions, indicating the encoder's ability to distinguish these surface types in the latent space.
  • Figure 3: Comparison of ground truth and model-predicted flood segmentation on the out-of-distribution Luxembourg flood event dataset. The visualization demonstrates that the LoRA approach generalizes better than the encoder-frozen approach to unseen geographical regions and flood events.