小样本条件下玉米单株干物质量无人机多光谱反演与模型稳定性分析

Estimation of Maize Single Plant Dry Matter Using UAV Multispectral Data under Small Sample Conditions: Model Stability Analysis

摘要:

为提升干旱区复播玉米单株尺度干物质量反演精度并支撑精准灌溉决策，针对小样本条件下模型稳定性不足的问题，以南疆复播玉米为研究对象，基于无人机多光谱影像开展单株干物质量预测研究。选取吐丝期、灌浆期和成熟期获取冠层影像，提取NDVI、GNDVI、NDRE、LCI和OSAVI等植被指数，并结合地面同步测定的株高与茎粗等结构变量，构建光谱-结构融合预测模型。在小样本条件下，采用岭回归（Ridge）、随机森林（RF）及集成模型进行建模，对不同变量组合及生育时期下模型精度与稳定性进行系统分析。结果表明：各植被指数与单株干物质量均呈显著正相关，其中NDRE与LCI表现最优，灌浆期相关性最高。结构变量能显著提升模型预测精度，且具有明显阶段性特征，吐丝期以株高主导，灌浆期和成熟期以茎粗贡献更为稳定。在灌浆期，模型预测效果最佳，Ridge模型验证集决定系数R2由0.71提升至0.90，集成模型由0.70提升至0.87，结构变量与光谱信息协同作用最为显著。单株尺度条件下模型精度对生育时期具有显著依赖性，整体由大到小依次为灌浆期、吐丝期、成熟期。研究结果表明，在小样本条件下，融合红边敏感植被指数与结构变量能够有效提升单株干物质量预测精度，灌浆期为最优建模时期。研究结果可为干旱区精准灌溉分区管理与水分调控提供数据支撑。

Abstract:

Aiming to improve the inversion accuracy of single-plant dry matter in relay-planted maize in arid regions and support precision irrigation decision-making, focusing on the insufficient stability of prediction models under small-sample conditions.Taking relay-planted maize in southern Xinjiang as the research object, single-plant dry matter prediction was conducted based on UAV multispectral imagery.Canopy images were acquired at the silking, grain-filling, and maturity stages, and vegetation indices, including NDVI, GNDVI, NDRE, LCI, and OSAVI were extracted.Meanwhile, structural variables such as plant height and stem diameter measured synchronously in the field were incorporated to construct spectral-structural fusion prediction models.Under small-sample conditions, ridge regression, random forest, and an ensemble model were used for modeling, and model accuracy and stability under different variable combinations and growth stages were systematically analyzed.The results showed that all vegetation indices were significantly positively correlated with single-plant dry matter, with NDRE and LCI performing best and the strongest correlations observed at the grain-filling stage.Structural variables improved model accuracy and showed clear stage-specific effects: plant height was more important at the silking stage, whereas stem diameter contributed more stably at the grain-filling and maturity stages.The best prediction performance was obtained at the grain-filling stage.The validation R2 of the ridge model was increased from 0.71 to 0.90, while that of the ensemble model increased from 0.70 to 0.87, indicating a strong synergistic effect between spectral and structural information.At the single-plant scale, model accuracy depended strongly on growth stage, with performance decreasing in the order of grain-filling, silking, and maturity stages.These findings suggested that integrating red-edge-sensitive vegetation indices with structural variables can effectively improve dry matter prediction under small-sample conditions, with the grain-filling stage being the optimal modeling period.The research results can provide data support for precision irrigation zoning and water regulation in arid regions.