Anisatul Latifah, Vivi Nur Wijayaningrum, Mamluatul Hani'Ah
Weather variability significantly impacts agricultural productivity, particularly in citrus farming, where crop growth is highly sensitive to climatic conditions. However, farmers often struggle to make timely decisions due to the lack of reliable weather-based crop management tools. To address this challenge, this study proposes a weather-based prediction and classification system designed to assist citrus farmers in identifying favorable and unfavorable weather conditions. The system utilizes daily weather data from the Meteorology, Climatology, and Geophysics Agency (BMKG) for Malang Regency from January 2022 to December 2024. Four key weather parameters, average temperature, relative humidity, rainfall, and sunlight duration were selected based on scientific literature and expert interviews. Long Short-Term Memory (LSTM) was applied to forecast each parameter, while Random Forest was used to classify the resulting conditions as 'favorable' or 'unfavorable.' The dataset was processed under four scenarios, namely standard preprocessing, augmentation, smoothing, and combining augmentation with smoothing. Experimental results showed that LSTM performed best with smoothed data, achieving RMSE values as low as 0.14 for some parameters. The Random Forest model achieved up to 99.5 % classification accuracy, and usability testing with citrus farmers resulted in a SUS score of 80.75, indicating strong user acceptance. The results suggest that applying machine learning to weather prediction and classification can significantly contribute to more informed and climate-resilient agricultural planning. © 2025 IEEE.
Politeknik Negeri Malang, Department of Information Technology, Malang, Indonesia