Soil salinization is the primary obstacle to the sustainable development of agriculture and eco-environment in arid regions. The accurate inversion of the major water-soluble salt ions in the soil using visible-near infrared (VIS-NIR) spectroscopy technique can enhance the effectiveness of saline soil management. However, the accuracy of spectral models of soil salt ions turns out to be affected by high dimensionality and noise information of spectral data. This study aims to improve the model accuracy by optimizing the spectral models based on the exploration of the sensitive spectral intervals of different salt ions. To this end, 120 soil samples were collected from Shahaoqu Irrigation Area in Inner Mongolia, China. After determining the raw reflectance spectrum and content of salt ions in the lab, the spectral data were pre-treated by standard normal variable (SNV). Subsequently the sensitive spectral intervals of each ion were selected using methods of gray correlation (GC), stepwise regression (SR) and variable importance in projection (VIP). Finally, the performance of both models of partial least squares regression (PLSR) and support vector regression (SVR) was investigated on the basis of the sensitive spectral intervals. The results indicated that the model accuracy based on the sensitive spectral intervals selected using different analytical methods turned out to be different: VIP was the highest, SR came next and GC was the lowest. The optimal inversion models of different ions were different. In general, both PLSR and SVR had achieved satisfactory model accuracy, but PLSR outperformed SVR in the forecasting effects. Great difference existed among the optimal inversion accuracy of different ions: the predicative accuracy of Ca²⁺, Na⁺, Cl⁻, Mg²⁺ and SO₄²⁻ was very high, that of CO₃²⁻ was high and K⁺ was relatively lower, but HCO₃⁻ failed to have any predicative power. These findings provide a new approach for the optimization of the spectral model of water-soluble salt ions and improvement of its predicative precision.

Soil salinization is a major problem of land degradation in arid and semiarid irrigation districts. This study aims to characterize the spatiotemporal evolution of soil salinization in Hetao Irrigation District (HID) in Inner Mongolia, China, using Landsat Thematic Mapper/Enhanced Thematic Mapper Plus/Operational Land Imager datasets. Salty barren land and farmland are extracted using supervised classification. Then, we develop four integrated soil salinity models (ISSMs) to quantify the intensity of saline farmland. ISSMs are generated through deriving the parameters (E_VI-SIs), which integrate enhanced vegetation index (EVI) and Salinity Index-1 (SI1), EVI and Salinity Index-3 (SI3), Modified Soil Adjusted Vegetation Index (MSAVI) and SI1, and MSAVI and SI3, respectively, from the scatter plots of farmland soils with different salinity in four spectral feature spaces (SFSs). Exponential regression analyses reveal that the E_VI-SI from MSAVI-SI3 SFS has the best fit with in situ soil electrical conductivity measurements (R² = 0.74, root mean square error = 0.31 dS m^–1). Salty barren land clustered in the central and northeast of HID, while the area of salty barren land decreased during 1986–2016. After employing water-saving irrigation since 2000, saline farmland decreased and then remained relatively stable. This study indicates that the SFS integrating MSAVI and SI3 contains effective information for quantifying the saline farmland. Employing water-saving irrigation had a positive effect on controlling salinization.

Accurate estimation of pan evaporation (E_p) is vital for the development of water resources and agricultural water management, especially in arid and semi-arid regions where it is restricted to set up the facilities and measure pan evaporation accurately and consistently. Besides, using pan evaporation estimating models and pan coefficient (kp) models is a classic method to assess the reference evapotranspiration (ET₀) which is indispensable to crop growth, irrigation scheduling, and economic assessment. This study estimated the potential of a novel hybrid machine learning model Coupling Bat algorithm (Bat) and Gradient boosting with categorical features support (CatBoost) for estimating daily pan evaporation in arid and semi-arid regions of northwest China. Two other commonly used algorithms including random forest (RF) and original CatBoost (CB) were also applied for comparison. The daily meteorological data for 12 years (2006–2017) from 45 weather stations in arid and semi-arid areas of China, including minimum and maximum air temperature (T_min, T_max), relative humidity (RH), wind speed (U), and global solar radiation (R_s), were utilized to feed the three models for exploring the ability in predicting pan evaporation. The results revealed that the new developed Bat-CB model (RMSE = 0.859–2.227 mm·d⁻¹; MAE = 0.540–1.328 mm·d⁻¹; NSE = 0.625–0.894; MAPE = 0.162–0.328) was superior to RF and CB. In addition, CB (RMSE = 0.897–2.754 mm·d⁻¹; MAE = 0.531–1.77 mm·d⁻¹; NSE = 0.147–0.869; MAPE = 0.161–0.421) slightly outperformed RF (RMSE = 1.005–3.604 mm·d⁻¹; MAE = 0.644–2.479 mm·d⁻¹; NSE = −1.242–0.894; MAPE = 0.176–0.686) which had poor ability to operate the erratic changes of pan evaporation. Furthermore, the improvement of Bat-CB was presented more comprehensively and obviously in the seasonal and spatial performance compared to CB and RF. Overall, Bat-CB has high accuracy, robust stability, and huge potential for E_p estimation in arid and semi-arid regions of northwest China and the applications of findings in this study have equal significance for adjacent countries.

Accurate estimation of the leaf area index (LAI) is essential for crop growth simulations and agricultural management. This study conducted a field experiment with rice and measured the LAI in different rice growth periods. The multispectral bands (B) including red edge (RE, 730 nm ± 16 nm), near-infrared (NIR, 840 nm ± 26 nm), green (560 nm ± 16 nm), red (650 nm ± 16 nm), blue (450 nm ± 16 nm), and visible light (RGB) were also obtained by an unmanned aerial vehicle (UAV) with multispectral sensors (DJI-P4M, SZ DJI Technology Co., Ltd.). Based on the bands, five vegetation indexes (VI) including Green Normalized Difference Vegetation Index (GNDVI), Leaf Chlorophyll Index (LCI), Normalized Difference Red Edge Index (NDRE), Normalized Difference Vegetation Index (NDVI), and Optimization Soil-Adjusted Vegetation Index (OSAVI) were calculated. The semi-empirical model (SEM), the random forest model (RF), and the Extreme Gradient Boosting model (XGBoost) were used to estimate rice LAI based on multispectral bands, VIs, and their combinations, respectively. The results indicated that the GNDVI had the highest accuracy in the SEM (R² = 0.78, RMSE = 0.77). For the single band, NIR had the highest accuracy in both RF (R² = 0.73, RMSE = 0.98) and XGBoost (R² = 0.77, RMSE = 0.88). Band combination of NIR + red improved the estimation accuracy in both RF (R² = 0.87, RMSE = 0.65) and XGBoost (R² = 0.88, RMSE = 0.63). NDRE and LCI were the first two single VIs for LAI estimation using both RF and XGBoost. However, putting more than one VI together could only increase the LAI estimation accuracy slightly. Meanwhile, the bands + VIs combinations could improve the accuracy in both RF and XGBoost. Our study recommended estimating rice LAI by a combination of red + NIR + OSAVI + NDVI + GNDVI + LCI + NDRE (2B + 5V) with XGBoost to obtain high accuracy and overcome the potential over-fitting issue (R² = 0.91, RMSE = 0.54).