Leveraging Deep Learning for Change Detection in Bi-Temporal Remote Sensing Imagery

Abstract

Deforestation in the Brazilian Amazon poses significant threats to global climate stability, biodiversity, and local communities. This dissertation presents advanced deep learning approaches to improve deforestation detection using bi-temporal Sentinel-2 satellite imagery. We developed a specialized dataset capturing deforestation events between 2020 and 2021 in key conservation units of the Amazon.

We first adapted transformer-based change detection models to the deforestation context, leveraging attention mechanisms to analyze spatial and temporal patterns. While these models showed high accuracy, limitations remained in effectively capturing subtle environmental changes.

To address this, we introduce DeforestNet, a novel deep learning framework that integrates advanced semantic segmentation encoders within a siamese architecture. DeforestNet employs cross-temporal interaction mechanisms and temporal fusion strategies to enhance the discrimination of true deforestation events from background noise. Experimental results demonstrate that DeforestNet outperforms existing models, achieving higher precision, recall, and F1-scores in deforestation detection. Additionally, it generalizes well to other change detection tasks, as evidenced by its performance on the LEVIR-CD urban building change detection dataset.

This research contributes a robust and efficient framework for accurate change detection in remote sensing imagery, offering valuable tools for environmental monitoring and aiding global efforts in sustainable forest management and conservation.