Multi-parametric multiple kernel deep neural network for crop yield prediction

Abstract

A rapid, precise and credible prediction of crop yield at a wider scale is more important than ever for crop management, the measurement of food production, food trading and policymaking to address the challenges of environmental change, increased population and food demand. Deep Learning (DL) models are recently well-known for predicting crop yields. Multi-parametric Deep Neural Network (MDNN) is a DL model employed to estimate the crop yield concerning multiple parameters such as climate and soil. A Growing-Degree Day (GDD) has been used to determine the impact of climate change on crop yield. The determined values along with the climate and soil parameters have been learned by the DNN to estimate the yield quality. The MDNN performs well with the huge volume of data and it is not suitable for the medium scale of data. The learning ability of MDNN is improved in this paper by proposing a Multi-parametric Multiple Kernel DNN (MMKDNN) to provide better crop yield prediction for medium-scale data. The effectiveness of the model built in the last hidden layer is solely determined by the intermediate representations of the input. The intermediate representation of the input in a neural network is combined through multiple kernel learning. The MMKDNN with increasing complexity representations is preserved in this way, but the output calculation, i.e. crop yield prediction, is free to use all of the network's knowledge. For five different types of crops, the experiments are conducted to assess the efficiency of the MMKDNN.