Abstract:
Soil salinization has seriously threatened food security and ecological stability in recent years, due to the degradation and crop yield reduction. It is often required to rapidly and accurately obtain soil salinization information in the sustainable production of farmland. One of the typical ' no irrigation, no agriculture ' areas is found in the Bianwan Farm of the Taolai River Basin in Gansu Province, China. There is less rain/drought and strong evaporation. The average annual precipitation and evaporation are 72.6 and 2184.0 mm, respectively. However, the conventional field sampling cannot monitor the large-scale soil salinization, due mainly to its time-consuming, labor-intensive, and costly nature. Remote sensing has emerged as an effective technical support in precision agriculture. This study aims to invert the soil salinity at the wheat tillering stage using UAV images and broad learning. Bianwan Farm was used as the representative sampling area of soil salinization. The multispectral of UAV remote sensing was used to capture the image data and the soil salt content in the wheat field in 2023 and 2024. The spectral reflectance and texture of the sampling points were then extracted from the remote sensing images. The red edge band was introduced to calculate the spectral index. The feature variables were optimized using Pearson correlation coefficient (PCC), grey relational analysis (GRA), and variable importance in projection (VIP). Scheme 1 (spectral index), Scheme 2 (texture feature), and Scheme 3 (spectral index-texture feature) were used as the model input groups. And then 54 models were constructed using Broad learning system (BLS), Back-Propagation Neural Network (BPNN), and Random Forest (RF). The spatial distribution of the soil salinity was then obtained according to the optimal model. A comparison was made on the inversion accuracy of the characteristic variables with the different models. The optimal inversion model of the soil salinity was determined for farmland in the irrigation area. The results show that the better inversion was achieved in the BLS model with Scheme 3 (spectral index-texture feature) as the input group among the different combinations of the characteristic variables. The coefficient of determination (
Rp2) on the validation set of the optimal model in 2023 was 0.851, the root mean square error (RMSEP) was 0.032%, and the average absolute error (MAEP) was 0.027%. The
Rp2 of the optimal model in 2024 was 0.811, RMSEP was 0.058%, and MAEP was 0.033%. Furthermore, the inversion accuracy of the BLS was better than that of the BPNN and RF models. The Rp2 values in 2023 were 0.189 and 0.289 higher than those of the BPNN and RF, respectively. The
Rp2 in 2024 increased by 0.161 and 0.145, respectively, indicating a better representation of the soil salt content. The BLS and three screening were achieved in the excellent inversion of the coupling model. The high robustness of the PCC-VIP-BLS coupling model was also obtained with the
Rp2/
Rc2 above 0.867. The finding can provide a strong reference and theoretical basis for the rapid inversion of the soil salt content in the saline-alkali soil in the arid area of northwest China.