Soil temperature prediction based on 1d-cnn-mlp neural network model

Soil temperature plays an important role in agriculture. In order to achieve cost reduction in the sensor ar-rangement when monitoring soil temperature, a novel model called 1D-CNN-MLP (One dimensional convolutional neural network-Multilayer perceptron) was proposed for soil temperature prediction. Meteorological data and soil temperature data on different soil layers collected for the 2018 similar to 2021 period from a weather station in Yangling, China, were used for calculation in our work. Our model was evaluated using statistical measures of MSE (Mean Square error). The model parameters with high operation efficiency and high accuracy are obtained, and the training result records much lower error than MLP (multilayer perceptron) and faster convergence than LSTM (long short-term memory) with an MSE of 0.288 x 10(-3). The 1D-CNN (One-dimensional convolutional neural network) part of the model is used to reveal and extrap-olate the law of how soil temperature propagates in different soil layers. In the case where only three layers of soil temper-ature data are known, the characteristic temperature layer depths of 10 cm, 15 cm, and 40 cm, are selected to place sensors and obtain the best prediction effect of soil temperature at different depths of 5 to 160 cm with a RMSE (Root mean squared error) of 1.988celcius. The model may help users with improved and economical soil temperature prediction and control, thus boosting crop yield. Ultimately, we found the model has a relatively poor performance in the accuracy of deep soil temper-ature prediction when only three layers of soil temperature data are known, and it is suggested that the model can be further optimized in terms of kernel parameter setting, data composition, and the variation law of deep soil temperature.