Digital assessment of banana (Musa spp.) genotype resistance to banana weevil (Cosmopolites sordidus) compared with expert visual assessment

Accurately assessing weevil damage is critical when evaluating banana germplasm. This enables identification of genotypes resistant to the banana weevil, Cosmopolites sordidus, for use as elite parents or for advancement in banana breeding programs. Although visual observation remains the most common phenotyping approach, it is limited by individual bias. This study investigated the potential of image analyses as precise and objective alternatives for assessing weevil damage on the banana corm. ImageJ and four machine learning models, to include YOLO_v11 medium, U-Net 512, SegFormer, and MiT-b0, were explored. 260, 65, and 72 images were used to train, test, and validate the machine learning models, respectively. Phenotyping trials were set up as partially replicated (P-rep) designs with 18 test genotypes and four control genotypes. All plants were produced through tissue culture, raised in pots before infestation with the weevils. At termination, the percentage score of each of the 370 corm samples was evaluated both visually and by image analysis to compare scores across all methods. A significant genotype effect was detected, indicating differences among genotypes in resistance to banana weevils. Furthermore, a significant genotype-by-scoring-method interaction showed that genotype rankings in terms of resistance to banana weevils varied across methods. This emphasized that the choice of scoring approach can affect the magnitude of damage quantified. Visual observation agreed more closely with image analyses for smaller scores and less for larger scores. Results from genotype performance evaluation showed that machine learning methods, except for the YOLO_v11, have a strong level of agreement and can be used interchangeably, giving consistent, reliable, and repeatable measurements. We recommend and have adopted machine learning when scoring weevil damage in the banana corm to avoid individual bias and subjectivity arising from visual observation.