以克拉玛依市春玉米叶片的SPAD值作为研究对象，利用无人机获取的多光谱遥感影像进行光谱反射率提取和植被指数的构建，并与SPAD值作相关性分析。分别以光谱反射率和相关性显著的植被指数作为输入变量，采用偏最小二乘法(PLS)、随机森林回归(RF)、萤火虫算法优化随机森林回归(FA-RF)、粒子群算法优化随机森林回归(PSO-RF)构建不同生长阶段的反演模型。结果表明，对于传统的线性回归模型，以光谱反射率为模型输入变量的模拟精度较优。对于机器学习模型，以植被指数为模型输入变量的模拟精度较优。对比4种模型综合评价指标，发现基于植被指数为输入变量的PSO-RF更适合反演春玉米不同时期的SPAD值，其拔节期、抽雄期、灌浆期验证集的R2值为0.914、0.901、0.928，RPD为2.355、2.543、2.655。

Using multispectral remote sensing images acquired by drones, spectral reflectance extraction and vegetation index construction were performed, followed by a correlation analysis with SPAD values. Spectral reflectance and vegetation indices with significant correlation were used as input variables to construct inversion models for different growth stages using Partial Least Squares(PLS), Random Forest Regression(RF), Firefly Algorithm-optimized Random Forest Regression(FA-RF), and Particle Swarm Optimization-optimized Random Forest Regression(PSO-RF). The results showed that for the traditional linear regression model, using band reflectance values as the model input variable yields better simulation accuracy. For the machine learning models, using vegetation indices as input variables provides better simulation accuracy. Comparing the comprehensive evaluation indicators of the four models, it was found that the PSO-RF model, with vegetation indices as input variables, is more suitable for inverting SPAD values of spring corn at different growth stages. The R2 values for the joint validation sets during the jointing, tasseling, and filling stages were 0.914, 0.901, and 0.928, respectively, with RPD values of 2.355, 2.543, and 2.655.

S513.01

国家自然基金项目(52169013)、新疆维吾尔自治区“十四五”重大专项(2020A01003-4)、自治区研究生科研创新项目(XJ2024G126)