Estimation of soil salt content by combining UAV-borne multispectral sensor and machine learning algorithms

PeerJ

Guangfei Wei¹  Junying Chen¹  Congcong Lao¹  Zhitao Zhang¹  Huifang Chen¹  Yu Li¹  Zhihua Yao¹  Yinwen Chen² 

[1] College of Water Resources and Architectural Engineering, Northwest A&F University, Yangling, China;Department of Foreign Languages, Northwest A&F University, Yangling, China;

关键词: Soil salt content;    Unmanned aerial vehicle (UAV);    Multispectral sensor;    Variable selection methods;    Machine learning algorithms;    Estimation models;   

DOI  :  10.7717/peerj.9087

来源: DOAJ

【 摘 要 】

Soil salinization is a global problem closely related to the sustainable development of social economy. Compared with frequently-used satellite-borne sensors, unmanned aerial vehicles (UAVs) equipped with multispectral sensors provide an opportunity to monitor soil salinization with on-demand high spatial and temporal resolution. This study aims to quantitatively estimate soil salt content (SSC) using UAV-borne multispectral imagery, and explore the deep mining of multispectral data. For this purpose, a total of 60 soil samples (0–20 cm) were collected from Shahaoqu Irrigation Area in Inner Mongolia, China. Meanwhile, from the UAV sensor we obtained the multispectral data, based on which 22 spectral covariates (6 spectral bands and 16 spectral indices) were constructed. The sensitive spectral covariates were selected by means of gray relational analysis (GRA), successive projections algorithm (SPA) and variable importance in projection (VIP), and from these selected covariates estimation models were built using back propagation neural network (BPNN) regression, support vector regression (SVR) and random forest (RF) regression, respectively. The performance of the models was assessed by coefficient of determination (R2), root mean squared error (RMSE) and ratio of performance to deviation (RPD). The results showed that the estimation accuracy of the models had been improved markedly using three variable selection methods, and VIP outperformed GRA and GRA outperformed SPA. However, the model accuracy with the three machine learning algorithms turned out to be significantly different: RF > SVR > BPNN. All the 12 SSC estimation models could be used to quantitatively estimate SSC (RPD > 1.4) while the VIP-RF model achieved the highest accuracy (Rc2 = 0.835, RP2 = 0.812, RPD = 2.299). The result of this study proved that UAV-borne multispectral sensor is a feasible instrument for SSC estimation, and provided a reference for further similar research.

【 授权许可】

Unknown