Laser & Optoelectronics Progress, Vol. 55, Issue 11, 113002 (2018)
Prediction of Soil Moisture Content by Selecting Spectral Characteristics Using Random Forest Method
Bao Qingling1,2, Ding Jianli1,2,*, and Wang Jingzhe1,2
- 1 Key Laboratory of Wisdom City and Environmental Modeling, College of Resource and Environment Sciences, Xinjiang University, Urumqi, Xinjiang 830046, China
- 2 Key Laboratory of Oasis Ecology, Ministry of Education, Xinjiang University, Urumqi, Xinjiang 830046, China
In order to more accurately analyze the importance of spectral absorption characteristic parameters, which in different soil moisture absorption bands in soil spectra, in soil moisture content estimation, we collect 38 soil samples in Ugan-Kuqa river oasis in Xinjiang to measure soil spectral reflectance and soil moisture content. The characteristic parameters of spectral water absorption are extracted with the continuum-removal method, the features include the maximum absorption depth D, the absorption peak right area Ra, the absorption peak left area La, the absorption peak total area A, area normalization maximum absorption depth DA, and symmetry S. With the correlation analysis of the features and soil moisture content, we use random forest method to classify the characteristic parameters of spectral water absorption, and obtain the importance of each parameter to soil moisture content. Multiple stepwise regression model is used to establish soil moisture content inversion model. The results are as follows: D and A have the strongest correlation with the soil moisture content, the correlation between spectral absorption parameters in the band of 2200 nm or 1400 nm and SMC is better than that of 1900 nm band; the top five parameters that are important for soil moisture content are obtained, they are D2200, La2200, A2200, D1900 and Ra2200, respectively; the best prediction model of SMC is the multiple stepwise regression model with A2200 and D2200, the decision coefficient of the modelling set is 0.88, root mean square error of modeling set is 2.08, decision coefficient of the test set is 0.89, prediction root mean square error is 2.21, and the relative analysis error is 2.80. Random forest classification can obtain the important spectral water characteristic parameters which have great influence on soil moisture content, and it provides a new method for accurate and rapid estimation of soil moisture content in arid areas.