Improving generalisability and transferability of machine-learning-based maize yield prediction model through domain adaptation

Abstract

Modern problems in agricultural management require non-traditional solutions, one of which is by utilizing domain adaptive machine learning models for crop yield prediction which are able to perform reliably in different temporal or spatial domains. However, most studies have focused on the application of domain adaptation to classification tasks such as crop type identification, while the application to regression tasks such as crop yield prediction have been limited. In this study, we explore the generalisability and transferability of ordinary Deep Neural Network (DNN) and domain adaptive neural network models created using three domain adaptation algorithms, namely Discriminative Adversarial Neural Network (DANN), Kullback-Leibler Importance Estimation Procedure (KLIEP), and Regular Transfer Neural Network (RTNN). These three algorithms represent feature-based, instance-based, and parameter-based domain adaptations, respectively. Maize yield records, weather variables, and remotely sensed features from 11 states in the US corn belt acquired in 2006–2020 were compiled and segregated into classes according to temporal (year) and spatial characteristics (annual growing degree days [GDD], vapor pressure deficit [VPD], soil organic content [SOC], and green chlorophyll vegetation index/GCI). We found that models trained using datasets from temperate regions with medium-high GDD and moderate VPD perform well whereas SOC does not significantly affect the generalisability. It is not advisable to train models with datasets constrained by GCI as this feature correlates significantly with the maize yield, and adaptation between two domains that rarely intercept will not work well. We also demonstrate that Kullback-Leibler divergence computed using features from source and target domains can be used to justify the feasibility of domain adaptation. Based on the divergence, a model trained in the US (or another region with sufficient data) is expected to work reliably in other regions through domain adaptation, especially feature-based adaptation.