Assessment for soil nutrient content prediction model based on visible-near infrared spectroscopy in the typical regions of Guangdong Province

Objective Visible-near-infrared spectroscopy (VNIRS) can non-invasively and rapidly predict soil nutrient contents based on models developed using data of some representative soil samples, but soil nutrient content prediction models for different regions of Guangdong Province are still lacking. In this study, representative agricultural soil samples were collected from Guangdong Province, and models were developed based on the traditional soil chemical analyses and VNIRS spectral signatures. The objective was to assess the predictive ability of the models for soil total and available nutrients so as to provide a scientific basis for rapid soil nutrient content estimation and soil quality evaluation in Guangdong Province.

Method A total of 514 representative soil samples were collected from the east (Meizhou), west (Zhanjiang), north (Shaoguan), and northwest regions (Zhaoqing) and the Pearl River Delta (Huizhou and Zhuhai) in Guangdong Province, and the contents of soil organic matter, total nitrogen, dissolved organic carbon, available nitrogen, and available phosphorus were analyzed. In addition, the VNIRS spectral signatures of soil samples were obtained between 400 and 2 490 nm. Models were developed and calibrated using partial least squares regression combined with principal component analysis, and the models were reversely validated and assessed for their predictive abilities.

Result There were significant differences in the contents of soil organic matter, total nitrogen, dissolved organic carbon, available nitrogen and available phosphorus and VNIRS spectral signatures among the soils from different regions. The models for soil organic matter and total nitrogen showed good prediction performances, with validation determination coefficients of 0.831 1 for soil organic matter in the northwest region and 0.789 8 for total nitrogen in the Pearl River Delta. The models for dissolved organic carbon, available nitrogen, and available phosphorus showed very different prediction performances among regions, with much better performances for available nitrogen and available phosphorus in the northwest region and Pearl River Delta compared with other regions. Model validation showed good correlations between the predicted and measured values of soil organic matter and total nitrogen, with the highest coefficients of determination (R²) being up to 0.69 and 0.65, respectively. The predicted and measured values of available nitrogen in the northwest region and Pearl River Delta also showed good correlations, with R² of 0.63 and 0.62, respectively. However, model validation showed generally poor correlations between the predicted and measured values of dissolved organic carbon and available phosphorus.

Conclusion VNIRS technology can predict soil organic matter and nutrient contents to differentiate soils from different regions at the provincial scale. Soil VNIRS spectral signature can be used as an important index for soil classification and soil quality evaluation. VNIRS technology shows good performance in the predictions of soil organic matter and total nitrogen contents, but its performance in the predictions of dissolved organic carbon, available nitrogen, and available phosphorus contents is element- and region-dependent. Therefore, future efforts should be focused on spectral range selection and model optimization.