Enhancing Weather Monitoring for Agriculture with Deep Learning: Anomaly Detection in East Java Using LSTM Autoencoder and OCSVM

Abstract

Agricultural productivity in East Java is under threat from unpredictable and harsh weather patterns, particularly rapid variations in sunlight length and rainfall intensity. These abnormalities can interrupt agricultural cycles, lower yields, and make farming communities more vulnerable to climatic calamities. However, current weather monitoring systems frequently fall short of detecting small anomalies in time series weather data that could serve as early warning signs of such disasters. This study seeks to close this gap by creating a robust anomaly detection methodology adapted to time-dependent weather variables important to agriculture. In this study, a hybrid model combining Long Short-Term Memory (LSTM) autoencoder and One-Class Support Vector Machine (OCSVM) is proposed. The LSTM autoencoder's structure reconstructs time series data and signifies anomalies through reconstruction errors (MSE), while OCSVM validates these anomalies to reduce false positives. The model was applied to daily weather data from East Java spanning 2015–2024. The results showed that the model effectively detected 11 anomalies in sunlight duration and 7 in rainfall, with F1-scores of 0.71 and 0.82, respectively. Several of these anomalies corresponded to actual disaster events such as floods, landslides, and droughts. This research contributed to the field by demonstrating the effectiveness of combining deep learning and machine learning for weather anomaly detection. The proposed framework offers valuable insights for early warning systems and can support local governments and farmers in improving disaster preparedness and enhancing agricultural resilience in East Java.