Toward Seasonal Wildfire Early Warning: Deep Learning-Based Forecasting of Active Fire Hotspots Using NDMI Time Series

Forest and vegetation fires are recurring environmental challenges in northern Thailand, contributing to seasonal haze, fine particulate matter (PM2.5) pollution, and wildfire-management difficulties. This study developed a province-level seasonal hotspot-forecasting framework for four fire-prone provinces: Chiang Mai, Mae Hong Son, Nan, and Uttaradit. Monthly active fire hotspot counts from NASA FIRMS MODIS, Normalized Difference Vegetation Index (NDVI), and Normalized Difference Moisture Index (NDMI) time series computed from MODIS MOD09A1 surface reflectance were aggregated at the provincial scale for the historical period from August 2017 to August 2025. Province-specific Long Short-Term Memory (LSTM) models were trained using 6-month input sequences of monthly NDMI together with temporal and seasonal encodings to predict the subsequent month hotspot count. To strengthen model defensibility, the revised framework incorporates baseline model comparison, walk-forward validation, ablation experiments, and observed validation of the forecast period where observed post-forecast hotspot data are available. Exploratory analysis showed clear seasonal hotspot dynamics, with hotspot activity increasing during the dry season and declining after the onset of rainfall. In contrast to NDVI, whose contemporaneous correlation with hotspot counts is weak, monthly NDMI showed a strong negative correlation with hotspot counts across provinces (Pearson r between -0.73 and -0.80), consistent with lower fuel moisture preceding higher fire activity. The initial LSTM test evaluation captured seasonal hotspot patterns with R2 values of 0.648 for Chiang Mai, 0.740 for Mae Hong Son, 0.685 for Nan, and 0.784 for Uttaradit; using NDMI improved test performance over NDVI in all four provinces. The 12-month forecast from September 2025 to August 2026 projected peak hotspot activity in March 2026, with 3,041 hotspots in Mae Hong Son, 2,619 in Chiang Mai, 1,348 in Nan, and 948 in Uttaradit. Observed validation across all four provinces using available FIRMS/GEE observations from October 2025 to April 2026 showed limited but positive explanatory skill, highest for Chiang Mai (R 2 = 0.478, WAPE = 69.70%), indicating that the model captured the seasonal escalation but underestimated peak hotspot magnitude. These findings suggest that LSTM-based hotspot forecasting can provide useful seasonal early-warning signals for wildfire preparedness, fuel management, and PM2.5 mitigation planning, although future operational deployment should integrate meteorological, fuel-moisture, topographic, and human-activity predictors.