Key points are not available for this paper at this time.
Wheat is a major global staple, and accurate yield estimation is critical for agricultural management and food security. While satellite and UAV-based approaches are widely used, they are costly and operationally complex, making ground-based RGB imagery an attractive low-cost alternative. However, existing image-based methods are often limited by the restricted representational power of handcrafted features or the limited interpretability of end-to-end deep learning models. This study therefore develops a ground-based RGB framework to systematically evaluate the roles of explicit handcrafted features and deep implicit features in wheat yield estimation. Explicit features describe spike morphology and canopy color using spatial geometry, wavelet-based texture, and visible vegetation indices, while implicit features are extracted from a VGG-based segmentation encoder. Four regression models (RF, SVM, KNN, and XGBoost) were tested. The results show that both feature types are informative, but implicit features provide superior predictive performance. In particular, XGBoost with implicit features achieved the highest accuracy (R 2 = 0.85, RMSE = 1015.77 kg·ha⁻ 1 , rRMSE = 12.9%), outperforming models based only on explicit features and matching the performance of combined feature sets. Overall, the proposed approach clarifies the relative roles of different feature representations and provides a flexible and practical solution for low-cost, scalable crop yield assessment based on ground-based RGB imagery.
Dong et al. (Fri,) studied this question.