Multi-Model Ensemble Depth Adaptive Deep Neural Network for Crop Yield Prediction

Abstract

Accurate prediction of crop yield enables critical tasks such as identifying the optimum crop profile for planting, assigning government resources and decision-making on imports and exports in more commercialized systems. In past few years, Machine Learning (ML) techniques have been widely used for crop yield prediction. Deep Neural Network (DNN) was introduced for crop yield. The crop yield prediction accuracy based on DNN was further improved by Multi-Model DNN (MME-DNN). It predicted the crop yield by modeling climatic, weather and soil parameters through statistical model and DNN. The MME-DNN is not scalable when new data appears consecutively in a stream form. In order to solve this problem, an Online Learning (OL) is introduced for crop yield prediction. In OL, DNN is learned in an online setting which optimizes the objective function regarding shallow model. But, a fixed depth of the network is used in ODL and it cannot be changed during the training process. So, Multi-Model Ensemble Depth Adaptive Deep Neural Network (MME-DADNN) is proposed in this paper to adaptively decide the depth of the network for crop yield prediction. A training scheme for OL is designed through a hedge back propagation. It automatically decides the depth of the DNN using Online Gradient Descent (OGD) in an online manner. Also, a smoothing parameter is introduced in OL to set a minimum weight for every depth of DNN and it also contributes a balance between exploitation and exploration. The crop yield is predicted from the soil, weather and climate parameters and their variation over four years by applying the MME-DADNN. Thus, by adaptively changing the depth of the DNN the performance of crop yield prediction is enhanced.