关键词: 连续小波变换, BP神经网络, 支持向量机, 精准农业

Abstract: Taking organic matter contents of 120 soil samples and corresponding spectral data in Tuoketuo County as data sources, the feasibilities of hyperspectral inversion of soil organic matters under different type of soils and lands  of different use were explored. The original spectrum (R), spectral reciprocal (1/R), spectral logarithm (LnR) and spectral first-order differential (R′) were decomposed by continuous wavelet transform to generate wavelet coefficients, and the correlation between soil organic matter and wavelet coefficients was analyzed, and BP neural network and support vector machine (SVM) were established by extracting the characteristic bands. The results were followed. ①The correlation coefficients between R, 1/R, LnR, R′ and soil organic matter were increased by 0.204, 0.090, 0.199 and 0.252 after continuous wavelet transform, respectively, which showed that continuous wavelet transform could deeply mine the potentially spectral information and enhance the correlation with organic matter content. ② Before continuous wavelet processing, SVM could not predict the content of soil organic matter, while after processing, the accuracies (R2) of SVM-CWT-R and SVM-CWT-R′ were 0.50 and 0.56, Root mean square errors (RMSE) were 0.17 and 0.15, residual predictive deviations (RPD) were 1.62 and 1.53, respectively, which realized the effective estimation of SOM. ③ After continuous wavelet transform, the results of BP neural network prediction model were improved. Among them, BP-CWT-LnR prediction model had the best effect, R2 was 0.76, which was higher than BP-lnR, RMSE was 0.15 reduced by 0.04, RPD was 2.12 increased by 0.87. Therefore, the BP-CWT-LnR hyperspectral inversion model could provide theoretical reference and technical support for precision agriculture.

Key words: continuous wavelet transform, BP neural network, support vector machine, precision agriculture