Point-by-point responses uploaded (R1 & R2); R3 response attached due to missing upload button

Linqiang DengYaoyu LiXifeng LiuZhimin ZhangJuanjuan MuShujie JiaYuqiao YanWuping Zhang^*

• Shanxi Agricultural University, Jinzhong, China

Frequent droughts and climate variability constrain stable yield increases in drought-tolerant crops. Focusing on sorghum in the Lifang dryland experimental area of Jinzhong, Shanxi, we developed a three-dimensional “spectral–meteorological–spatial” yield-prediction framework. Multispectral images (16 vegetation indices) were acquired with a DJI Mavic 3M at the emergence, jointing, flowering, and maturity stages, while daily meteorological data were collected simultaneously. Eight machine-learning algorithms were compared, and SHAP was used to quantify variable contributions. Ensemble methods performed best: Gradient Boosting and Random Forest achieved R² values of 0.9491 and 0.9070, respectively, markedly outperforming linear models. The most informative features were DVI and NDGI, followed by TVI, SAVI, and GRVI. Growth stage strongly affected prediction, with the jointing stage contributing the most (R² = 0.9454), followed by maturity and flowering, and the emergence stage contributing least (R² = 0.1371). The optimal model produced a yield map of 4,291–4,965 kg·ha⁻¹ with a distinct banded pattern; Moran’s I = 0.5552 indicated moderate positive spatial autocorrelation characterized by “high–high” and “low–low” clusters. This work provides technical support for precision planting and efficient sorghum management in arid regions.