Soil salinity inversion at different depths using improved spectral index with UAV multispectral remote sensing

Abstract: Quick and accurate acquisition of soil salinity information with vegetation cover is critical to prevent environmental deterioration especially in arid and semi-arid areas. The UAV multispectral remote sensing systems has widely been expected to apply for monitoring the soil salinity, due to its low cost, high resolution, as well as resistance to weather and terrain. This study aims to obtain the soil salinity at various depths under the crop cover, using the improved spectral index. The UAV multispectral remote sensing images were captured at four test sites with different salinization degrees, including 0.065%-0.275%, 0.194%-0.828%, 0.220%-1.239%, 0.594%-3.112%, in Shahaoqu Irrigation Area, Inner Mongolia, China (40°52′-41°00′N, 107°05′-107°10′E, elevation 1030 m), from July 16 to 20 in 2019. Simultaneously, the soil salinity data were collected with various depths at 0-10 cm, 10-20 cm, and 20-40 cm. Firstly, a six-rotor UAV equipped with a Micro-MCA multispectral camera was used to acquire the images, where the traditional spectral index was calculated using the extracted spectral reflectance with remote sensing images. A Rededge band based on the traditional spectral index was introduced to establish a new spectral index, serving as an improved spectral index. Next, an Elastic-net algorithm (ENET) was selected such spectral variables as spectral band, traditional spectral index, and modified spectral index (established by introducing Rededge band). The screened spectral variables were divided into the original spectral variable group and the improved spectral variable group. Finally, three machine learning algorithms, such as BP Neural Network (BPNN), Support Vector Machine (SVM), and Extreme Learning Machine (ELM), were combined with the ENET to construct the soil salinity inversion model at different soil depths. The maps of soil salt inversion were drawn at the test sites using the optimal inversion model that constructed by the improved spectral variable group, after evaluating the model performance. The results showed that: 1) The variable selection method ENET can be used to effectively screen the optimal spectral variables. The performance of inversion models that constructed by three variable selection methods was superior to those without screening variables; 2) The optimal inversion depth of soil salinity with vegetation cover was >10-20 cm. The model performance of ELM was better than that of SVM and BPNN. The ENET-ELM inversion model performed better, where the determination coefficients (RC2) of calibration dataset were 0.785, the root mean square error (RMSEC) were 0.128, the consistency correlation coefficients (CC1) were 0.879, the determination coefficients (RP2) of validation dataset were 0.783, the root mean square error (RMSEP) were 0.141, and the consistency correlation coefficients (CC2) were 0.875. 3) At different soil depths, the soil salinity inversion map that drawn by the optimal inversion model using the improved spectral variables can effectively elucidate the degree of salinization in the test area, indicating that the introduction of Rededge band to construct the spectral index can be used for the soil inversion of salinity. This finding can provide a promising way for using UAV multi-spectral remote sensing to monitor and prevent soil salinization of farmland.