Improving model robustness for soybean iron deficiency chlorosis rating by unsupervised pre-training on unmanned aircraft system derived images

Abstract Iron deficiency chlorosis (IDC) is a major yield-limiting factor for soybean production in the mid-western USA. The most practical solution in mitigating losses due to IDC is the development and characterization of IDC tolerant varieties. Leveraging the advanced technique of unmanned aircraft system (UAS) and the thriving deep learning methodology, a convolutional neural network (CNN) could be trained to assist breeders with IDC resistance selection. However, a known difficulty in IDC screening is that the symptoms often vary across diverse genetic backgrounds and spatial or temporal soil heterogeneities. A robust CNN model is desired to mitigate such difficulty. While high robustness usually relies on a sufficiently large labeled training data, the available labeled samples in most breeding programs are normally not enough. Under this limitation, it is critical to find an alternative way to train a robust model. The solution proposed in this study was to apply unsupervised pre-training on the unlabeled aerial images that are much easier to obtain by the UAS. Specifically, a convolutional autoencoder (CAE) was pre-trained on unlabeled sub-images clipped from aerial RGB images; then, the pre-trained weights were reused to initialize the CNN model that was trained on labeled plot-wise sub-images clipped from stitched RGB maps. To test the robustness of this CAE initialized model (CAE1-CNN), two baseline models were equally trained: the first was CAE2-CNN, where the CAE2 was pre-trained with three times of unlabeled data as that of CAE1, by adding wniter wheat and sorghum aerial images; the second was Ran-CNN where the CNN was randomly initialized. Three conditions were considered for testing model robustness: different soybean trials, field locations and vegetative growth stages. Results revealed that both the CAE1-CNN and the CAE2-CNN had relatively better robustness than the Ran-CNN model, i.e., higher R2 and lower RMSE values, especially on different soybean trials and growth stages, which proved that the unsupervised pre-training added gains to the model robustness across diverse trials and growth stages. Similar performances were found between the CAE1-CNN andthe CAE2-CNN model, suggesting that augmenting the unlabled data did not bring significant improvement to model robustness. Additionally, during robustness test on different soybean trials, the unsupervised pre-training seemly showed the potential of alleviating the required number of labeled training samples. These promising findings could contribute to the research on crop stresses by providing a potential path towards developing a robust system for classifying or predicting stress severities under more varied conditions.