Dual-Stream Deep Learning for Deforestation Content-Based Image Retrieval with Optical and Infrared Satellite Imagery

Abstract

The speed and scale of forest degradation in Indonesia demands sophisticated environmental data journalism and high-fidelity visual storytelling tools to communicate ecological problems effectively to the public and policy makers. Nevertheless, current satellite-based monitoring approaches that rely solely on unisensory optical data are not sufficient for discriminating visually similar tropical land cover types (e.g., natural forests vs. industrial plantations), thus limiting the reliability of visual evidence in environmental media campaigns. Furthermore, the extreme class disparities in real-world satellite datasets limit the ability of automated algorithms to preserve important aspects of minority degradation patterns. In this research, we present a dual-stream deep learning architecture that fuses optical (RGB) and infrared (IR) modalities for better descriptive power of geographic visual communication utilizing a late-fusion technique. To prevent the statistical bias to the major classes, we propose a Label Smoothing regularization and a four-way Test-Time Augmentation (TTA) methodology for visual inference to guarantee the geometric invariance and prediction stability. We compared the dual-stream SwinV2 transformer with eight state-of-the-art architectures and it surpasses the convolutional baselines with an overall accuracy of 80.54% and a greater macro F1-Score on extremely unbalanced datasets. Importantly, the recovered latent embeddings allow a powerful Content-Based Image Retrieval (CBIR) engine. This framework is an intelligent visual asset retrieval system that allows media designers, journalists and environmental campaigners to dynamically explore and develop captivating visual narratives of deforestation trends across the Indonesian archipelago.