Integrate deep learning and physically-based models for multi-step-ahead microclimate forecasting

Abstract

Precision agriculture control systems count on reliable and accurate microclimate forecasts to maintain envi-ronmental suitability for crop growth. However, IoT devices adopted to monitor microclimate are expensive to people in developing countries. This study proposed a hybrid deep learning model (ConvLSTM*CNN-BP) without using IoT data to produce accurate multi-horizon and multi-factor (greenhouse internal temperature, relative humidity, and photosynthetically active radiation) forecasts simultaneously. The proposed model fused a convolutional-based long short term memory neural network (ConvLSTM), a convolutional neural network (CNN), and a backpropagation neural network (BPNN). Model construction involved an ensemble of gridded 6 -hour-ahead meteorological forecasts from the STMAS-WRF model and 3-hour-ahead greenhouse internal tem-perature simulated by a physically-based model at a 10-min scale. Another deep learning model (CNN*LSTM*Stacked LSTM-BP) using IoT data was established for comparison purpose. The experimental results on two greenhouses located in Central Taiwan indicated that the proposed model (non-IoT) and the benchmark model (with IoT) produced similar forecast performances on greenhouse internal temperature, relative humidity, and photosynthetically active radiation. Moreover, the proposed model with illustrious abilities of noise removal and feature extraction could provide satisfactory forecast accuracy. The proposed deep learning approach hits a milestone in multi-horizon and multi-factor forecasting on microclimate, which significantly supports farmers, especially in developing countries, in reducing the installation and maintenance costs of IoT devices for moni-toring purpose.