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

•Soybean IDC scores can be predicted by the convolutional neural network (CNN)•Varied IDC symptoms across diverse soil heterogeneities require a robust CNN model.•Unlabeled UAS derived RGB images can be leveraged via unsupervised pre-training.•Pre-training enhanced CNN robustness on different soybean trials and growth stages.•Pre-training could probably alleviate required number of labeled training samples.