Deep Learning Architecture for Tomato Plant Leaf Detection in Images Captured in Complex Outdoor Environments

Abstract

The detection of plant constituents is a crucial issue in precision agriculture, as monitoring these enables the automatic analysis of factors such as growth rate, health status, and crop yield. Tomatoes (Solanum sp.) are an economically and nutritionally important crop in Mexico and worldwide, which is why automatic monitoring of these plants is of great interest. Detecting leaves on images of outdoor tomato plants is challenging due to the significant variability in the visual appearance of leaves. Factors like overlapping leaves, variations in lighting, and environmental conditions further complicate the task of detection. This paper proposes modifications to the Yolov11n architecture to improve the detection of tomato leaves in images of complex outdoor environments by incorporating attention modules, transformers, and WIoUv3 loss for bounding box regression. The results show that our proposal led to a 26.75% decrease in the number of parameters and a 7.94% decrease in the number of FLOPs compared with the original version of Yolov11n. Our proposed model outperformed Yolov11n and Yolov12n architectures in recall, F1-measure, and mAP@50 metrics.

Code and data availability

The authors explicitly state that the data (custom tomato leaf detection dataset with ground-truth annotations) and code used in this paper are publicly available in their GitHub repository andros1206/Leaf-Detection. This is a paper-specific, publicly actionable asset reproducing the paper's phenotyping images/labels (