Using computer vision to identify seed-borne fungi and other targets associated with common bean seeds based on red–green–blue spectral data

The science of seed pathology has been established since the development and application of standardized methods for assessing seed health to meet the needs of the seed industry and associated regulatory entities. Despite seed health testing being a routine operation in most countries, results of testing often vary from one laboratory to another. We evaluated computer vision using red–green–blue (RGB) imagery and machine learning algorithms to detect seed-borne fungi on common bean ( Phaseolus vulgaris L.) seeds. Seeds of common bean were submitted to the standard blotter test for 7 days, followed by fungal identification using a stereo- and light microscope. A scanning electron microscope was used to confirm fungal identity. Images of seed-borne fungi were captured from a distance of approximately 5 cm. Seventeen spectral indices were derived from the RGB images. Targets of interest in the images were obtained using spatial polygons with attributes used for training six machine learning algorithms (random forest (rf), rpart, rpart1SE, rpart2, naive Bayes, and svmLinear2), with a total of five replicates per target that were identified as Aspergillus flavus , A. niger , A. ochraceus , Penicillium sp., Mucor sp, Rhizopus sp, Fusarium sp, Rhizoctonia sp., common bean tegument, and blotter paper. After a fivefold cross-validation process and a confusion matrix, the rf algorithm had the highest prediction success to detect the targets (accuracy 0.80 and Kappa 0.77, respectively). The brightness index was the most important variable in predicting targets by the rf. Using the rpart1SE algorithm, a decision tree for target identification was obtained with an accuracy of 0.70 and a Kappa value of 0.66, respectively. The rf, svmLinear2, and rpart1SE were found to be the most robust classification algorithms for predicting identification of the fungal species and other targets associated with common bean seed blotter tests using digital RGB images and indices. The use of spectral indices derived from RGB imagery has extended the training capability of algorithms, demonstrated by the importance of the variables and decision tree used for target prediction by the rf and rpart1SE algorithms, respectively.