Deep-learning-based wildfire occurrence prediction with human factors in Gangwon Province, South Korea

The occurrence of wildfires and their damage have increased due to desiccation caused by global warming. Numerous studies have attempted to predict the occurrence of wildfires. However, predicting wildfires remains challenging owing to unpredictable anthropogenic activities and fluctuating environmental conditions. This study predicts the likelihood of wildfire occurrence in Gangwon Province, South Korea, by considering all relevant factors: meteorology, topography, and anthropogenic activities. A deep learning model based on a multimodal convolutional neural network was developed to incorporate nonlinear factors including temperature, humidity, wind speed, precipitation, normalized difference vegetation index, elevation, slope, population density, and land surface. The training dataset consisted of 2106 historical cases, 1053 wildfire and 1053 non-wildfire occurrences between 2011 and 2021. According to the results of classification metrics, the model performance improved from 69 % when using only meteorological variables to 74 % when including topographic data together with the environmental index NDVI, and further to 77 % when incorporating all data, including human factors. The error map analysis reveals no systematic regional bias and spatial autocorrelation in model outcomes. A combination-based model evaluation and a partial dependence plot (PDP) analysis were conducted to assess the contribution of each variable. Analysis of 31 combinations of input features revealed that incorporating human-related factors improved the model's performance. Furthermore, PDP analysis revealed a significant association between land use—particularly in the residential, industrial, and agricultural sectors—and the increased likelihood of wildfires. In a practical test, our model, with all factors, more effectively identified high-risk areas in the 2022 Yanggu wildfire. This study suggests that a multimodal approach effectively integrates social and environmental factors, thereby improving the accuracy of wildfire prediction.