Comparison of two inversion methods for winter wheat leaf area index based on hyperspectral remote sensing

Abstract: Leaf area index (LAI) is an important index for evaluating winter wheat's growth status and forecasting its yield. Hyperspectral remote sensing is a new technical approach that can be used to acquire the information of winter wheat LAI immediately. By integrating hyperspectral remote sensing and traditional field monitoring, this study aims to explore the best simulation accuracy and adaptability to the different method of high spectral monitoring winter wheat leaf area index inversion. Two kinds of universal LAI inversion methods based on hyperspectral remote sensing data through regression analysis method and the BP neural network (BPNN) are introduced and used in this study. The study area is Changqing district of Jinan city, Shandong province, China's Huang-huai-hai plain. On winter wheat growth stage, the winter wheat canopy spectral reflectance and LAI were monitored in field using the ASD FieldSpec 3 and SunScan canopy analysis system. The study selected the following 6 vegetation index (RVI, DVI, NDVI, GRVI, EVI and SAVI) combined with spectral reflectance characteristics of the study area. The 6 vegetation indexes are closely related to winter wheat LAI with correlation at a significant level. After correlation analysis of the Hyperspectral Vegetation Index (HVI) and LAI, winter wheat LAI regression models and BPNN model were established. Then simulation precisions for different models were analyzed and evaluated. The 6 winter wheat LAI regression models fits were 0.696~0.775, and root mean square errors (RMSE) were 0.386-0.523. Accuracy test showed that NDVI inversion model had the highest accuracy compared to other models. It is concluded that NDVI model is the most suitable model for inverting winter wheat LAI in the study area. However, the NDVI inversion model must avoid saturation phenomenon when NDVI is close to 1. This is the model's inadequacy. Input multiple sensitive reflectivity bands contain 450, 550, 670 and 870 nm bands to the BP neural network model. Upon examination, the simulation and measured fit values was 0.990 and the RMSE was 0.105. The results show that BP neural network model inversion method can build a better LAI inversion for winter wheat varieties in different regions. Among them, the inversion model has the highest R2 (0.990) and least RMSE (0.105). The BP neural network method used to construct the inversion model is better on different varieties of winter wheat LAI inversion. However, establishing BP model needs to ensure enough samples (generally the number of samples n>50 is a large sample of events) for the research adaptability. Both methods have their advantages and disadvantages. Overall, inversion method should be selected according to the number of samples and monitoring area.