Optimization of LSTM and Linear Regression Models with Hyperparameters and Genetic Algorithm for Forest and Land Fire Risk Prediction

Abstract

Forest and land fires are a serious environmental challenge in Indonesia, especially in peatland areas such as Kubu Raya Regency, West Kalimantan. This study aims to develop and compare two forest and land fire risk prediction models, namely LSTM and Linear Regression, each optimized using hyperparameters and Genetic Algorithms. Climate data (temperature, rainfall, humidity, wind speed) and hotspots from 2019–2024 were used as input features. The pre-processing stage included normalization and handling of missing data. Forest and land fire risk was categorized into three classes based on thresholds set by Manggala Agni. The LSTM model was tested with one and two layers of neurons. The best results were obtained from the two-layer neuron architecture with an RMSE of 0.2364, MAE of 0.1955, and MAPE of 12.90%. As a comparison, the optimized linear regression model showed lower performance with an RMSE of 0.3026, MAE of 0.2602, and MAPE of 17.95%. These results indicate that LSTM is superior in recognizing temporal and non-linear patterns in time series data. Spatial analysis indicates that the districts of Sungai Raya, Sungai Kakap, and Kuala Mandor B are high-risk areas, particularly toward the end of the dry season. These findings underscore the potential of LSTM-based deep learning approaches in supporting early warning systems and mitigation strategies for forest and land fires. Integrating predictive models with spatial analysis could serve as an effective strategy for more targeted forest and land fire management.