Evaluation Of CNN-LSTM with Attention for Forest Fire Prediction in Indonesia: Challenges of Imbalanced Data

Abstract

Forest and land fires are annual disasters in Indonesia that are influenced by the temporal and nonlinear dynamics of surface weather conditions. This study evaluated a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture with an attention mechanism for predicting forest fire risk based on multivariate surface weather data from BMKG stations for the 2015–2024 period. The dataset comprised 4,931 daily observations with eight weather features and four fire risk levels. The models were evaluated using temporal cross-validation and compared with the baselines (Random Forest, XGBoost, LightGBM, SVM, and standalone LSTM). The results indicate that the CNN-LSTM-attention model achieved 71.47% accuracy and 62.36% F1-score in the initial configuration. However, optimization attempts using SMOTE substantially degraded the performance to 50.95% accuracy (F1=0.55). The critical findings are as follows: (1) SMOTE is unsuitable for time-series data as it disrupts temporal patterns; (2) simpler architectures outperform excessive complexity; (3) class weighting is preferable to oversampling for handling class imbalance; and (4) basic weather features (without extensive engineering) yield optimal results. This study concludes that deep learning approaches for forest fire prediction in tropical regions face significant challenges related to class imbalance and temporal stability, necessitating the development of more adaptive modeling strategies