Pest-induced damage detection in corn leaves using improved Faster R-CNN

Pest-induced damage detection in cereal crops, such as corn, is challenging because of the small and irregular nature of damage symptoms, which can closely resemble plant diseases or environmental stress and foreign-attached substances in leaves. Delays in identification or misdiagnosis can result in substantial yield and economic losses. Previous studies have primarily focused on pest species or diseases. This study presents a task-specific adaptation of Faster R-CNN for detecting pest-induced damage in corn leaves caused by Grasshoppers and Fall Armyworms by customising the RPN with parallel dilated convolutions and an adaptive attention module. A new field-collected corn leaf pest-induced damage dataset comprising 1,390 images, was annotated in collaboration with an agricultural expert. Forty images containing the artefact were embedded in the training dataset to enhance the model performance. The model was assessed using five-fold cross-validation, and four major evaluation metrics were evaluated. The introduction of the artefacts into the training data resulted in an improvement in mAP, along with a reduction in misdetections attributable to the artefacts. Furthermore, the incorporation of parallel dilated convolutions in the RPN of Faster R-CNN led to enhancements in mAP 50:95 (54.06%) and exhibited a high equilibrium between precision and recall, achieving an F1-score of 80.13% with increased computational cost. Compared with the baseline Faster R-CNN model, the proposed model achieved a +1.38-percentage point increase in recall for Fall Armyworm damage and a +1.79-percentage point improvement in mAP 50:95 for Grasshopper damage. This result suggests an improved detection of small objects (Fall Armyworms) and accurate localisation of larger, irregular objects (Grasshopper), both individually and when co-occurring. This research suggests a potential future applicability for farmers in the timely identification of pest infestations, thereby mitigating yield losses. These findings provide guidance for future research on detecting a wider variety of pest-induced crop damages across diverse geographical regions in agricultural computer vision applications.