Hyperspectral Feature Extraction and Estimation of Soil Total Nitrogen Based on Discrete Wavelet Transform

Soil total nitrogen is an important nutrient index. Hyperspectral technology is used to study and build a hyperspectral estimation model of total nitrogen content in Shajiang black soil, which provides a reference for crop fertilization and the development of precision agriculture. This paper attempts to study the feasibility of discrete wavelets to estimate soil total nitrogen content. Taking different wheat nitrogen fertilizer treatments in Shangshui County, Henan Province, as the experimental area, 100 samples of Shajiang black soil with a depth of 0 similar to 20 cm were collected. After the soil samples were air-dried in the dark and processed by grinding and screening, the spectra were collected in the darkroom of the laboratory. The total samples (100 sand ginger black soil) were divided into 75 modeling sets and 25 validation sets. The original spectrum was transformed by the first derivative, and the first derivative spectrum was analyzed by correlation analysis and discrete wavelet transform respectively. At the same time, the hyperspectral estimation model of soil total nitrogen content was constructed by combining the support vector machine and the k-nearest neighbor algorithm. The correlation between the single band of the original spectrum and the first derivative spectrum and soil total nitrogen were systematically analyzed. The results showed that after the first derivative transformation, the spectrum had a better correlation with soil total nitrogen, and the correlation coefficient reached 0.84 at 1 373 nm. The discrete wavelet algorithm selects the best mother wavelet and decomposition level of the first derivative spectrum. The results show that the wavelet coefficients decomposed by the Sym8 function can better reconstruct the spectral information of soil total nitrogen. Further, based on the low-frequency coefficients of decomposition layer L-1-L-11, the support vector regression and k-nearest neighbor regression estimation models of soil total nitrogen content were established respectively, and all the estimation models were compared. The model constructed by combining the low-frequency coefficients of decomposition layer L-5 with k-nearest neighbor is the best. The determination coefficient of modeling is 0.90, the root mean square deviation is 0.09 g.kg(-1), and the relative analysis error is 3.78. The validation determination coefficient is 0.97, the root mean square deviation is 0.05 g.kg(-1), and the relative analysis deviation is 4.30. At the same time, compared with the model constructed with the full band and the sensitive band selected after correlation analysis as input, the accuracy of the K-neighbor model is improved by 3.2% and 9%, and the accuracy of the support vector machine model is improved by 6.7% and 11.6%. The results show that the first derivative transform and discrete wavelet technology can effectively suppress the impact of noise, improve the estimation accuracy of soil total nitrogen content, reduce the dimension of spectral data, simplify the complexity of the model, and provide a reference for the accurate estimation of the total nitrogen content of Shajiang black soil.