Improving localized weather predictions for precision agriculture: A Time-Series Mixer approach for hazardous event detection

Abstract

Natural environmental systems and human activities are deeply interconnected, especially in agriculture. Despite advancements in agricultural techniques, weather remains a critical factor influencing crop yields and livestock health. Precision agriculture relies on weather predictions to mitigate environmental risks caused by weather. However, numerical weather predictions are generated by global or regional numerical models, lacking the resolution to capture site-specific conditions. Artificial intelligence can address this gap by integrating numerical weather predictions data with local station observations. This study employs the Time-Series Mixer (TSMixer) neural network to forecast temperature, wind speed, relative humidity, and precipitation over a 45-hour horizon. Trained with predictions from the MOLOCH model and data from ARPA stations near six agricultural sites in Northern Italy, TSMixer achieves greater accuracy than the MOLOCH model. Additionally, TSMixer excels in detecting hazardous events for precision agriculture, including frost damage, heat stress, and germination block, highlighting its value for environmental risk management.