土壤水分含量(SMC)的快速估测对干旱半干旱地区的精准农业发展具有重要的意义。 以渭干河-库车河绿洲为靶区, 采用小波变换(WT)对反射光谱进行1~8层小波分解, 通过相关性分析确定最大分解层数, 再通过竞争性自适应重加权(CRAS)、 连续投影算法(SPA)和CARS-SPA耦合算法进行特征波长筛选。 基于全波段构建BP神经网络模型和基于特征波长构建BP神经网络、 支持向量机、 随机森林和极限学习机模型, 并进行对比分析。 结果显示: (1)随着小波分解的进行, 总体上L6在去噪的同时还尽可能的保留了光谱原始特征, 为最大分解层; (2)小波变换和CARS-SPA算法的结合使其在建立模型时较为彻底的去除噪声和无信息变量, 同时消除变量间的共线性; (3)在所有的SMC预测模型中, 相对于BP神经网络、 SVM, ELM和RF具有更好的预测能力, 其中L6-CARS-SPA-ELM精度最高, 其RMSEC=0.015 1, R2c=0.916 6, RMSEP=0.014 2, R2p=0.935 4, RPD=2.323 9。 这体现出ELM预测模型对非线性问题的强解析能力和模型的稳健性, 为该研究区SMC的预测提供新的思路。

The rapid estimation of soil moisture content (SMC) is of great significance to precision agriculture in arid areas, and hyperspectral remote sensing technology had been widely used in the estimation of soil moisture content due to its non-destructive, rapid, and high spectral resolution characteristics. Meanwhile, there are many prediction models of soil moisture content, such as BP, SVM, RF and so on, but the prediction model has some shortcomings. Recently, the extreme learning machine(ELM) as a new algorithm began to emerge in the field of soil property prediction. In the present study, a total of 39 soil samples at 0～20 cm depth were collected from delta oasis in Weigan-Kuqain, Xinjiang Province. We brought back to the laboratory to dry it naturally, groundnd and passed through a 2 mm hole scree, and then the sample holders were clear black boxs in 12 cm diameter and 1.8 cm deep, which were filled and leveled at the rim with a spatula. Reflectance of soil samples were measured using ASD Fieldspec 3 Spectrometer in a dark room. We used the following steps to process soil reflectance: First, discrete wavelet transformation (DWT) was used to decompose the original spectral in 8 levels using db4 wavelet basis by MATLAB programming language. In order to select the maximum level of DWT, correlation coefficients between SMC and the spectra of each level was computed. Secondly, On the basis of wavelet transform, CARS (the adaptive variable weighting algorithm), SPA (successive projections algorithm) and CARS-SPA were used to filter the redundant variables, the wavelength variables with better correlation with SMC were screened out. Thirdly, On the basis of the preferred wavelengths, BP neural network, SVM (support vector machine), RF (random forest) and ELM (extreme learning machine) prediction models were employed to build the hyperspectral estimation models of SMC, and the advantages and disadvantages of the model were further analyzed. Statistical parameters of root mean square error of calibration (RMSEC), determination coefficient of calibration (R2c), root mean square error of prediction (RMSEP), determination coefficient of predicting (R2p) and relative prediction deviation (RPD) were selected as comparison criteria. The results showed that: (1) With the increase of the number of decomposed layers, the correlation between soil reflectance and SMC showed a trend of increasing first and then decreasing, and L6 was the most significant band at 0.01 level. In general, the characteristic spectrum of L6 was denoised at the same time, and the spectral detail was preserved to the maximum extent. So the maximum decomposition order of the wavelet was 6 order decomposition; (2) On the basis of L6, the CARS, SPA and CARS-SPA algorithms were used to optimize the variables, and the number of selected wavelength variables were 81, 23 and 12, respectively. The predictive models constructed by three algorithms were better than those of the whole-band model. The prediction model based on the CARS-SPA was the most accurate in the corresponding model. It can be seen that the CARS-SPA coupling algorithm not only simplified the model complexity, but also increased the robustness of the model; (3) Compared with the BP, SVM, RF and ELM, In all the SMC predicting models, there were 6 models with predictive ability, Sort by: L6-CARS-SPA-ELM>L6-CARS-SPA-RF>L6-CARS-ELM>L6-CARS-RF>L6-SPA-ELM>L6-SPA-RF. Results showed that ELM performed much better than BP, SVM and RF in predicting SMC in this study. At the same time, the L6-CARS-SPA-ELM model had the highest accuracy, and the model had RMSEC=0.015 1, R2c=0.916 6, RMSEP=0.014 2, R2p=0.935 4, RPD=2.323 9. It was shown that the combination of wavelet transform and CARS-SPA algorithm made it possible to remove the noise as much as possible and to remove the noise completely when the model was established. At the same time, and ELM model was a new method to predict other soil properties.