关键词: 遥感数据, 森林可燃物含水率, 光谱水分指数, 深度学习

Abstract:

Objective: This paper was carried out to study the water content measurement based on satellite remote sensing data and to compare the accuracy of deep learning model and traditional machine learning model in order to provide a theoretical basis for the establishment of forest fuel water content database in China. Method: Taking Chongli District, Zhangjiakou City, Hebei Province as the research area, based on the field measurement data, aiming at the problem of a large error of traditional machine learning model, the mutilayer perceptron(MLP) model in deep learning was established. The relationships between remote sensing spectral reflectance and water content of forest canopy vegetation and litter were studied, and the accuracy was compared with the support vector regression(SVR) model in traditional machine learning. In the experiment, sentry remote sensing data of the same season as field investigation were selected, and then the model variables commonly used in remote sensing estimation method such as spectral reflectance and spectral moisture index were used as the influence factors of retrieving the moisture content of canopy vegetation and surface litter. Finally, the model training was carried out based on the data of field investigation. Inversion using the remote sensing estimation method for ever encountered in the process of surface litter moisture content of canopy cover problem, this study used in the processing of remote sensing data bidirectional reflectance distribution function to obtain samples of the remote sensing data of different observation angles, combined with the radiative transfer model, the canopy reflectance model mapped to the surface reflectance after training. Result: In the experimental result, the fitting degree of MLP model with red light, green light, near-infrared and short-wave infrared bands as input variables in the inversion of canopy vegetation water content was 0.843, which was better than the fitting degree of the optimal model in SVR of 0.807, and the accuracy improved by 4.5%. The fitting degree of MLP model in the inversion of surface litter water content was 0.448, which was better than the optimal model in SVR of 0.408, and the accuracy improved by 9.8%.In this study, the optimal fitting model MLP was used to invert the distribution map of fuel water content, and it was concluded that the water content of canopy vegetation was higher in the western Chongli district, while the water content of surface litter was higher in the southeastern Chongli district. Conclusion: The research result of this paper could demonstrate an optimization scheme to solve the problem of poor penetrability of optical remote sensing from canopy to surface, and could also provide a theoretical basis for using remote sensing estimation methods to measure moisture content of regional canopy vegetation and surface litter at a large scale.

Key words: remote sensing data, moisture content of forest combustibles, spectral moisture index, deep learning