ROMJIST Volume 29, No. 3, 2026, pp. 237-248, DOI: 10.59277/ROMJIST.2026.3.03
Palanichamy NAVEEN, Mahmoud HASSABALLAH A Spectral-Attention Deep Learning-Based Approach to Deforestation Monitoring in Satellite Imagery
ABSTRACT: In automated AI systems for urban change detection, deforestation tracking requires effective, scalable approaches that generalize across different ecosystems. Current Normalized Difference Vegetation Index (NDVI)-based approaches and single-biome deep learning networks lack good generalizability across regions and are energy-intensive. NDVI is prone to saturation over closed vegetation and fails over dry canopy; biome-specific networks overfit and require extensive labeled datasets. These limitations hinder accurate monitoring in diverse environmental settings. To this end, a spectral-attention U-Net-based model is introduced using Short-Wave Infrared (SWIR) bands and multi-temporal Landsat 8/9/Sentinel-2 imagery for biome-flexible deforestation monitoring with transfer learned ResNet-50 and carbon-friendly training regimes. SWIR improves vegetation distress surveillance and is less susceptible to atmospheric perturbations; transfer learning reduces data demands and training energy with sustainability improvements. Also, Spectral Attention Gates (SAGs) are proposed that dynamically weight spectral inputs for improved feature extraction, and integrate a lightweight decoder design using transposed convolution for sharper boundary detection. The proposed model achieves 92.3% IoU (Intersection over Union) in the Amazon and 88.7% in Jordan’s temperate forests with up to 22% higher precision in arid regions compared to the NDVI approach, while maintaining low energy costs (1.43 kgCO2e/10 km2 on NVIDIA A100). Hotspot analysis reveals 42.3 km2/year loss in the Amazon (Z=4.71) and 12.5 km2/year in Jordan (Z=3.28), validated against 5000 ground truth patches (κ = 0.82).KEYWORDS: Deep learning; deforestation detection; spectral attention; U-Net architecture; urban change detectionRead full text (pdf)