Deep learning algorithms enable accurate identification of cassava varieties (Manihot esculenta Crantz) using image...

Deep learning is an emerging, powerful technology that facilitates object detection in plant science, enabling applications such as disease diagnostics and identification of plant species and crop varieties using image analysis. In this study, we demonstrated the success of deep learning algorithms when used in identifying Tanzania-released cassava varieties (TARICASS4, Kiroba, and Mkuranga1) through image analysis. Three deep learning algorithms (YOLO v8, SSD-MobileNet v2, and Faster R-CNN with ResNet-50) were trained using a dataset of 19,800 images from leaves, petioles and stems of cassava plants of the three varieties. Trained algorithms achieved promising performance in identifying the cassava varieties using each of the three plant parts (mAP50≥85.7%), with the highest values recorded for leaves (mAP50≥92.7%). Of the three models, the YOLO v8 model showed the consistent promising performance for the three plant parts, with an mAP50 of 97.3% for leaves. Further findings from confusion matrices confirmed the superior performance of YOLO v8 for leaves, with performance of ≥98% for each of the three varieties. YOLO v8 also achieved high levels of performance for stems and petioles for all varieties. Our results confirm that deep learning models can be effectively applied for AI-based identification of field-grown cassava. The approach developed for cassava has strong potential for scaling to additional cassava varieties as well as other Vegetatively Propagated Crops, and for use by seed producers and regulators who could benefit from the use of a field-usable tool for verifying variety 'true-to-typeness'.