Objective Accurately acquire and analyze watermelon fruit phenotype information to support the breeding of new watermelon varieties.

Methods For different phenotypic features of watermelon, a combination of image processing, deep learning, and 3D reconstruction techniques was employed to automatically extract and analyze phenotypic traits such as fruit peel thickness, pulp area, pulp color, fruit length and width, cross-sectional seed count, and texture percentage.

Results The proposed method effectively extracts watermelon phenotypic traits, with no significant difference between image measurements and manual measurements (P>0.05). The average error RMSPE is <0.03, MAPE is <0.03, and the coefficient of determination R² is all greater than 0.94. The seed count recognition mAP value is 0.936.

Conclusion The method proposed in this study can rapidly, conveniently, and accurately extract watermelon fruit phenotypic traits, addressing current issues such as incomplete acquisition of watermelon phenotypic information, insufficiently accurate data, and excessive reliance on empirical judgment. This technology will contribute to the automation of watermelon breeding and germplasm resource evaluation, providing reliable data support for the breeding of new varieties. Moreover, it holds reference value for the analysis and application of fruit phenotypes in other crops.