A Carbon-Aware Explainable Transformer for Multi-Temporal Deforestation Detection: Balancing Accuracy and Environmental Impact

Abstract

Deep learning has significantly improved the performance of earth observation, including deforestation monitoring and land-use change detection. But, with the advent of modern deep neural networks, an extensive amount of computational power is needed, which consequently creates high energy consumption and carbon emissions. This poses a significant challenge for sustainable AI, especially in environmental monitoring systems aimed at facilitating climate action. Most current remote sensing techniques based on transformers are primarily geared towards enhancing the accuracy of the detection, with little consideration for computational sustainability and model interpretability. To address this, we propose a Carbon-Aware Explainable Transformer (CAE-Former) for multi-temporal deforestation mapping that balances accuracy, explainability, and efficiency. The approach consists of three main components: 1) real-time tracking of energy consumption and CO2 emissions during training, 2) a carbon-aware adaptive training scheme that dynamically adjusts batch size under a predefined carbon budget, and 3) an explainable temporal-spatial attention for model interpretability. The approach is tested on several benchmark datasets, showing robust results under various scenarios. The proposed framework achieves competitive performance across multiple remote-sensing benchmarks while maintaining an ultra-low carbon footprint. On the ONERA-CD dataset, CAE-Former attains an F1-score of 81.14% and an IoU of 68.26% with only 2.88 g CO2 emissions. Under the normalized carbon-intensity setting of 0.233 kg CO2/kWh, the reported emissions for the primary change-detection benchmarks remain below 3 g CO2, while India-grid-equivalent values are also reported using the Central Electricity Authority emission factor. These outcomes demonstrate the feasibility of incorporating carbon-aware optimization in deep learning processes to produce sustainable and explainable artificial intelligence for large-scale forest monitoring. The framework offers an avenue to sustainable AI in Earth observation.