What type of study is this?

This is a Quantitative Study study.

September 16, 2025Open Access

Deep learning-based approach for phenotypic trait extraction and computation of tomato under varying water stress

WYWeiyue YangChinese Academy of Agricultural Engineering JLJinshan LiKunming University of Science and Technology YFYayang FengFarmland Irrigation Research Institute

Key Points

The deep learning model improved tomato phenotype recognition, achieving a 5.4% increase in mAP50–95.
Key phenotypic parameters showed a relative error of 6.9% for plant height, indicating reliable analysis.
Random Forest led classification with up to 98% accuracy under various water stress conditions.
The approach supports early identification of water stress, benefiting precision irrigation practices.

Abstract

Introduction With the advancement of imaging technologies, the efficiency of acquiring plant phenotypic information has significantly improved. The integration of deep learning has further enhanced the automatic recognition of plant structures and the accuracy of phenotypic parameter extraction. To enable efficient monitoring of tomato water stress, this study developed a deep learning-based framework for phenotypic trait extraction and parameter computation, applied to tomato images collected under varying water stress conditions. Methods Based on the You Only Look Once version 11 nano (YOLOv11n) object detection model, adaptive kernel convolution (AKConv) was integrated into the backbone’s C3 module with kernel size 2 convolution (C3k2), and a recalibration feature pyramid detection head based on the P2 layer was designed. Results and discussion Results showed that the improved model achieved a 4.1% increase in recall, a 2.7% increase in mAP50, and a 5.4% increase in mAP50–95 for tomato phenotype recognition. Using the bounding box information extracted by the model, key phenotype parameters were further calculated through geometric analysis. The average relative error for plant height was 6.9%, and the error in petiole count was 10.12%, indicating good applicability and accuracy for non-destructive crop phenotype analysis. Based on these extracted traits, multiple sets of weighted combinations were constructed as input features for classification. Seven classification algorithms—Logistic Regression, Support Vector Machine, Random Forest, Decision Tree, K-Nearest Neighbors, Naive Bayes, and Gradient Boosting—were used to differentiate tomato plants under different water stress conditions. The results showed that Random Forest consistently performed the best across all combinations, with the highest classification accuracy reaching 98%. This integrated approach provides a novel approach and technical support for the early identification of water stress and the advancement of precision irrigation.

Ask AI

Helpful

Bookmark

View Full Paper