基于“珠海一号”的水稻高光谱定量估产研究

Rice Yield Estimation Based on OHS-1 Hyperspectral Remote Sensing Images

利用“珠海一号”高光谱卫星遥感数据实现水稻定量估产反演。首先抽稀地面实测光谱,与卫星数据光谱波段保持一致;其次通过皮尔逊相关分析,从光谱、光谱一阶导数以及NDVI、RVI、DVI、EVI2、VI、SAVI等6个植被指数中,计算并选取产量敏感性较高的18个建模因子;再次通过训练随机森林模型,获得最优训练结果下的最优参数模型;最后利用珠海一号卫星数据进行定量反演,得到研究区的水稻估产值。结果表明:基于4参数的多生育期随机森林估产模型精度较高,抽穗期光谱704 nm波段一阶导,灌浆期567 nm波段一阶导,灌浆期884 nm波段一阶导以及灌浆期DVI(λ_(1)=658 nm,λ_(2)=523 nm)为自变量的模型效果最佳;珠海一号水稻反演效果较好,反演得到铁桥村水稻平均产量575.2 kg/亩,与实际采集情况相差仅15.5 kg/亩。

We focused on rice yield estimation based on OHS-1 hyperspectral remote sensing images.Firstly,we diluted the measured spectra on the ground and maintain consistency with the spectral bands of satellite data.Then,we calculated the eighteen modeling factors by Pirsson correlation analysis,from the spectra,the first derivative of spectra and six vegetation indices(NDVI,RVI,DVI,EVI2,VI and SAVI),and trained the random forest model to obtain the optimal parameter.Finally,we carried out the quantitative inversion by OHS-1to obtain the rice yield.The results show that the rice yield estimation of multi-growth period random forest based on four parameters has high accuracy,with the first derivative at 704 nm at heading stage and 567 nm at filling stage,the first-order conduction at 884 nm and DVI at grain filling stage(λ_(1)=658 nm,λ_(2)=523 nm)is the best independent variables.The average yield of TieqiaoVillage rice is 575.2 kg/mu,with the only 15.5 kg/mu difference from the truth.