Deep Learning Models Using Multi-Modal Remote Sensing for Prediction of Maize Yield in Plant Breeding Experiments

Accurate grain yield predictions, as well as knowledge of environmental inputs and the temporal response of the crop through the growing season are important in maize plant breeding experiments. Advances in remote sensing technologies for providing measurements of plant characteristics over extended spatial areas have made high-throughput phenotyping increasingly accurate and much more affordable, particularly when coupled with advances in machine learning models. While data driven models in agriculture have become increasingly popular, the transferability between genotypes, environments, and their interactions is a relatively new domain. In this study, a temporal attention model based on Long Short-Term Memory (LSTM) networks is proposed for maize grain yield prediction. The model architecture leverages soil categorical variables and includes time-series of remote sensing and time-dependent weather variables. Stratified sampling for transfer learning was investigated to refine the pre-trained models. Results indicate that the proposed deep learning models can achieve high accuracies for individual year predictions. Furthermore, the proposed transfer learning technique with limited training provided high accuracies with reduced computational times compared to training the model from scratch.