Spectral features extraction for estimation of soil total nitrogen content based on modified ant colony optimization algorithm

Abstract

Nondestructive and rapid estimation of soil total nitrogen (TN) content by using near-infrared spectroscopy plays a crucial role in agriculture. The obtained original spectrum, however, presents several disadvantages, such as high redundancy, large computation, and complex model, because it generally processes a large amount of data. This study aimed to determine soil TN content-sensitive wavebands with high information quality, considerable predictive ability, and low redundancy. This paper proposes an evaluation criterion in selecting sensitive wavebands based on three factors, namely, degree of relevance with target variables, representative ability of the entire spectral information, and redundancy of the selected wavebands. Based on these three factors, two methods, namely, mutual information (MI) algorithm and the combination of ant colony optimization (ACO) and MI, were innovatively developed to identify soil TN content-sensitive wavebands. After the analysis and comparison, a set of wavelengths, including 943, 1004, 1097, 1351, 1550, 1710, 2123, and 2254 nm, using the ACO–MI combined method was selected as the soil TN content-sensitive wavebands to estimate the TN content of soil samples, under four soil types, collected from different regions. The partial least squares (PLS) models based on full-spectral information, multiple linear regression (MLR) models and support vector machine (SVM) regression models based on the eight selected wavelengths for soil TN content were established separately. After the comparison, the MLR and SVM models achieved higher accuracies than the PLS models based on the full spectral information. In addition, the SVM models got the best results. In the calibration group, the coefficients of determination (R2) was 0.989, and the root mean square errors (RMSE) of calibration was 0.078 g/kg. In the validation group, the R2 was 0.96, and the RMSE of prediction was 0.219 g/kg. The residual predictive deviation (RPD) was 5.426. For the soil samples with TN content in the range of 0–1 g/kg, the detection precision also reached a high level. Therefore, the eight sensitive wavebands selected through the ACO–MI method performed good mechanism, universality and predictive ability in soil TN content estimation. The ACO–MI method would be valuable for soil sensing in precision agriculture.