?
Mapping Forest Fire Risks Using Deep Learning: A Case Study in An Giang Province, Vietnam
With the escalating frequency and intensity of global wildfires, there is an urgent need for predictive models that balance accuracy with operational practicality. Building upon the TFDeepNN wildfire prediction framework of Truong et al. [1], this study develops a deep learning framework for wildfire risk assessment in Vietnam’s An Giang province, focusing on early detection of significant fire events. Building upon the TensorFlow-based deep neural network approach previously applied in Vietnam, we implement deployment-oriented improvements including dynamic fire growth metrics, higher-temporal-resolution MERRA-2 weather data integration, and improved JAXA LULC classifications. Our enhanced model achieves 91.4% overall accuracy with 74.8% recall for significant fires and 92.5% precision, while maintaining 97.7% specificity and only a 2.3% false alarm rate. At the same time, a recall of 74.8% means that 25.2% of significant fires are missed, so the model is best used as a decision-support layer alongside existing monitoring and verification procedures rather than as a standalone alarm source. Through feature importance analysis using SHAP, we identify fire spread rate as the dominant predictor (37.6% importance), supporting the physical plausibility of the model’s decisions. Overall, the contribution is best described as applied integration and optimization of multi-source data and deployment-aware evaluation, rather than a new deep learning method.