A multi‐task deep learning–based approach to identify and assess dry root rot disease in chickpea

Premise Dry root rot (DRR), an emerging global disease exacerbated by drought, affects a wide range of crops. The disease symptoms are manually scored for large‐scale germplasm screening to identify disease‐resistant varieties and test transgenic lines and other varieties. Despite its use in the past, manual disease identification is not only cumbersome but also often incapable of distinguishing between similar stress phenotypes. Manual disease scoring is tedious, time‐consuming, and often error‐prone, especially when screening large numbers of plants. To overcome this limitation, we developed a machine learning model that employs a learning‐based approach using convolutional neural networks for automated feature extraction and image‐based identification and severity assessment of DRR. Methods Our models were trained, tested, and validated using an in‐house dataset of 10,800 unique RGB images. A total of 51 deep learning models, including single‐task and multi‐task models, were developed using the Xception and Vision Transformer (ViT) architectures with standard image augmentations, and then validated on stratified independent test sets to demonstrate robustness across camera, scanner, and microscope modalities. Results The best multi‐task ViT model, capable of scoring across combined datasets, achieved 94% accuracy in distinguishing DRR from the control and had a root mean square error of 1 in predicting DRR severity score. Discussion This model's metrics, including confusion matrix performance, underscore its potential as an efficient and accessible tool for large‐scale germplasm screening for DRR, overcoming the challenges of manual scoring.