A Lightweight Deep Learning Framework for Long-Term Weather Forecasting in Olive Precision Agriculture

Mohamed Abdelwahab, Hassan Mostafa, Ahmed K. F. Khattab

Abstract: In this paper, a lightweight deep learning-based time series forecasting model is developed to predict the daily temperature values for one year ahead. The predictive model is an encoder-decoder model with a single LSTM layer for each of the encoder and decoder. Unlike the existing literature of time series forecasting, the proposed framework is designed to be lightweight to be deployed on low-complexity hardware platforms installed in the olive groves. Using real-life data of a Spanish olive grove, we show that the accuracy loss of the proposed lightweight framework is insignificant (0.004% to 0.06%). On the other hand, the implementation complexity of the proposed model is orders of magnitude lower than existing models, making it more suitable for implementation on embedded hardware platforms.