Geographia Technica, Vol 21(2), Special Issue: Artificial Intelligence Applications in Geography, 2026, pp. 35-50
PREDICTING FIRE HOTSPOTS IN KALIMANTAN BASED ON CLIMATE FACTORS USING GAUSSIAN PROCESS REGRESSION AND LONG SHORT-TERM MEMORY
Sri NURDIATI 
, Mochamad Tito JULIANTO , Ionel HAIDU , Muhammad Daryl FAUZAN , Hari NURDIANTO , Syukri Arif RAFHIDA , Mohamad Khoirun NAJIB
ABSTRACT: Forest and land fires in Kalimantan present a recurrent environmental challenge, driven by local and global climatic factors. Predicting fire hotspots is crucial for mitigation efforts. This study compares the performance of Gaussian Process Regression (GPR) and Long Short-Term Memory (LSTM) networks in forecasting monthly fire hotspots based on six climatic indicators, including rainfall, dry spells, and ENSO and IOD indices. GPR models were developed using several kernels and hyperparameter tuning methods, while LSTM models applied multiple architectural configurations combined with regularisation techniques. The results show that while GPR models achieved good fitting on training data, they suffered from overfitting and lower accuracy during testing, even after optimisation. In contrast, the LSTM model with two LSTM layers and four dense layers demonstrated superior predictive performance, achieving a testing RMSE of 522.12 and an Explained Variance Score (EVS) of 0.834. LSTM effectively captured complex temporal patterns inherent in climate-driven fire hotspot data. Nevertheless, both models faced difficulties in predicting anomalies linked to socio-economic interventions, such as the significant reduction in fire hotspots in 2018.The findings highlight the effectiveness of LSTM in modelling temporally dependent environmental phenomena and suggest the need for integrating socio-economic variables into future predictive frameworks to improve robustness. This study contributes valuable insights towards enhancing early warning systems for forest fire risk management in Kalimantan and other tropical regions.
Keywords: Fire hotspots; Kalimantan; Gaussian Process Regression; Long Short-Term Memory; Machine learning.
Full article here