区域冬小麦单产统计数据降尺度制图方法

Downscaling Mapping and Driving Factor Analysis of Regional Winter Wheat Yield Statistical Data

摘要:

区域冬小麦单产空间分布是作物精细化管理的重要基础,但受限于实测样本稀缺,精细制图仍面临挑战。农业统计数据连续稳定且权威,但其向像元尺度单产的降尺度应用不足,如何在保持统计一致性的同时刻画空间异质性,是亟待解决的问题。以中国北方冬小麦主产区为研究对象,提出一种区域冬小麦单产统计数据降尺度制图方法。首先,综合植被指数、气象、土壤及地形数据,利用多元线性回归(Multiple linear regression, MLR)、随机森林(Random forest, RF)和轻量级梯度提升机(Light gradient boosting machine, LightGBM)3 类回归模型,构建县域尺度单产模型;其次,在保持统计一致性的前提下,引入相对单产权重,将县级平均单产推至像元尺度,生成冬小麦单产精细化空间分布数据;最后,利用野外实测样点对降尺度结果进行精度验证,并基于SHAP(SHapley additive exPlanations)可解释性框架解析单产空间变化的驱动因子。结果表明:LightGBM 模型在县域尺度估产表现最优(R2为0.79,RMSE为581.51kg/hm2,MAE为401.48kg/hm2, MRE为8.20%,NRMSE为9.72% ),降尺度像元单产与实测数据具有较好一致性(R2为0.65,RMSE为623.32kg/hm2,MAE为540.65kg/hm2,MRE为8.45%,NRMSE为9.76%);研究区冬小麦降尺度像元单产介于943.96~9453.06kg/hm2之间,以中高产水平为主导的双峰分布特征,呈现东高西低且平原高、山地低的格局;SHAP可解释性分析发现,海拔、土壤粉粒含量、全磷含量、坡度和水分胁迫是研究区冬小麦单产高低变化的主导驱动因子,低产区受地形与土壤结构性约束显著,高产区则以养分因子为主要限制。本研究所提冬小麦单产统计数据降尺度方法可为单产统计数据降尺度提供方法参考,相关结果可为作物精细化管理、冬小麦精准种植决策提供数据支撑。

Abstract:

Spatial distribution of winter wheat yield is fundamental for precision crop management. However, fine-scale mapping remains challenging due to the scarcity of field-observed yield samples. Agricultural statistical data are continuous, stable and authoritative, but their application in pixel-scale yield downscaling is still limited. A key challenge is how to characterize spatial heterogeneity while maintaining statistical consistency. The main winter wheat producing area in northern China was taken as the research object, and a downscaling mapping method for regional winter wheat yield statistics was proposed. Firstly, based on the data of vegetation index, meteorology, soil and terrain, a county-scale yield model was constructed by using multiple linear regression (MLR), random forest (RF), and light gradient boosting machine (LightGBM). Secondly, on the premise of maintaining statistical consistency, the relative yield weight was introduced to push the average yield at the county level to the pixel scale, and the refined spatial distribution data of winter wheat yield per unit area were generated. Finally, the accuracy of the downscaling results was verified by field measured samples, and the driving factors of spatial variation of yield per unit area were analyzed based on SHapley additive exPlanations (SHAP) framework. The results showed that the LightGBM model performed best at the county scale (R2=0.79, RMSE=581.51kg/hm2, MAE=401.48kg/hm2, MRE=8.20%, NRMSE=9.72%), and the downscaled pixel yield was in good agreement with the measured data (R2=0.65, RMSE=623.32kg/hm2, MAE=540.65kg/hm2, MRE=8.45%, NRMSE=9.76%). The yield of winter wheat ranged from 943.96kg/hm2 to 9453.06kg/hm2, which was dominated by medium-high yield level and showed a bimodal distribution, showing a spatial distribution of high in the east and low in the west, high in the plain and low in the mountain. SHAP-based interpretability analysis further indicated that elevation, soil silt content, total phosphorus content, slope and water stress were the main driving factors affecting winter wheat yield variation. Low-yield areas were significantly constrained by topography and soil structure, while high-yield areas were mainly limited by nutrient factors. The proposed downscaling method can provide a useful reference for statistical yield downscaling and can support precision crop management and winter wheat cultivation decision-making.