Integrating Explainable Deep Learning with Multi-Temporal Land Cover Change Detection for Carbon Stock Estimation

High-density carbon stocks are being severely undermined by rapid urbanization and deforestation in the Western Ghats of Southern India. However, current monitoring systems operate at coarse spatial resolution, lack temporal currency, and offer no spectral interpretability, making credible carbon accounting and proactive conservation intervention impossible under current frameworks. This study suggests an integrated pipeline that applies explainable deep learning semantic segmentation, multi-temporal Sentinel-2 and Landsat 8/9 imagery, post-classification change detection, and GEDI- calibrated InVEST carbon stock estimation to 320,000 km² of Southern India between 2020 and 2026. The first spectral early-warning threshold for forest carbon loss is NDVI < 0.48, with impacted pixels 3.8 times more likely to undergo conversion, according to transition-level SHAP analysis. Three architectures, such as U-Net, DeepLabV3+, and Vision Transformer, optimized via Bayesian search and weighted ensemble combination, achieve 92.4% overall accuracy. With a net carbon loss of 131.8 Mt C (483.7 Mt CO2e, USD 87.3 billion) over six years, change detection indicates 18,340 km² of landscape transition led by urban growth (+6,840 km²) devouring Dense Forest (−3,170 km²). A transparent, repeatable, and spatially explicit decision-support tool directly relevant to India's 2030 NDC carbon monitoring, Western Ghats conservation planning, and REDD+ verification frameworks, the generated Carbon Risk Map reveals 14,050 km² at immediate conversion risk.