Soil salinization is a significant global ecological issue that leads to soil degradation and is recognized as one of the primary factors hindering the sustainable development of irrigated farmlands and deserts. The integration of remote sensing (RS) and machine learning algorithms is increasingly employed to deliver cost-effective, time-efficient, spatially resolved, accurately mapped, and uncertainty-quantified soil salinity information. We reviewed articles published between January 2016 and December 2023 on remote sensing-based soil salinity prediction and synthesized the latest research advancements in terms of innovation points, data, methodologies, variable importance, global soil salinity trends, current challenges, and potential future research directions. Our observations indicate that the innovations in this field focus on detection depth, iterations of data conversion methods, and the application of newly developed sensors. Statistical analysis reveals that Landsat is the most frequently utilized sensor in these studies. Furthermore, the application of deep learning algorithms remains underexplored. The ranking of soil salinity prediction accuracy across the various study areas is as follows: lake wetland (R² = 0.81) > oasis (R² = 0.76) > coastal zone (R² = 0.74) > farmland (R² = 0.71). We also examined the relationship between metadata and prediction accuracy: (1) Validation accuracy, sample size, number of variables, and mean sample salinity exhibited some correlation with modeling accuracy, while sampling depth, variable type, sampling time, and maximum salinity did not influence modeling accuracy. (2) Across a broad range of scales, large sample sizes may lead to error accumulation, which is associated with the geographic diversity of the study area. (3) The inclusion of additional environmental variables does not necessarily enhance modeling accuracy. (4) Modeling accuracy improves when the mean salinity of the study area exceeds 30 dS/m. Topography, vegetation, and temperature are relatively significant environmental covariates. Over the past 30 years, the global area affected by soil salinity has been increasing. To further enhance prediction accuracy, we provide several suggestions for the challenges and directions for future research. While remote sensing is not the sole solution, it provides unique advantages for soil salinity-related studies at both regional and global scales.

Current digital soil mapping of soil properties (soil organic carbon, SOC; electrical conductivity, EC; and pH) is mainly based on transfer learning, which is inadequate in terms of accuracy for the northern plain area of Xinjiang. To address this issue, establishing a new model is urgently required that can improve our understanding of the soil properties in this region. To this end, based on the global bioclimatic variables and surface dry–wet and wet–dry transitions, The study developed a spectral–water–heat database (SWHD). The study then incorporated this database and background data into machine learning algorithms (XGBoost, LightGBM, and random forest) to establish models applicable to the study area and draw spatial changes in the key soil properties. Our findings revealed that the organic carbon content was the highest in grasslands, whereas shrublands had high soil salinity. The pH value indicated overall alkalinity in the study area. Additionally, the SWHD-based predictions outperformed the mean or maximum value datasets, with LightGBM showing superior performance among all models. Furthermore, the validation accuracy obtained through our optimal algorithm was significantly higher than that obtained by other products, such as Harmonized World Soil Database (HWSD) and SoilGrid250, likely because of the limitations of these datasets, which may represent historical soil properties rather than current variations in the soil properties in the region. The study also observed that the mean SOC and EC values significantly decreased compared to the historical data, while the decrease in pH was smaller but not significant. Structural equation modeling and variable importance analysis revealed that the variables with the greatest influence on modeling SOC, EC, and pH were BIO10, DTW2021_406-426_B3 (Surface reflectance acquired in spring), and land use type. Our improved model developed based on the SWHD dataset offers important scientific evidence and decision support for land use management and provides a solid foundation for future research in this field.

Groundwater level (GWL) in dry areas is an important parameter for understanding groundwater resources and environmental sustainability. Remote sensing data combined with machine learning algorithms have become one of the important tools for groundwater level modeling. However, the effectiveness of the above-based model in the plains of the arid zone remains underexplored. Fortunately, soil salinity and soil moisture may provide an optimized solution for GWL prediction based on the application of apparent conductivity (ECa, mS/m) using electromagnetic induction (EMI). This has not been attempted in previous studies in oases in arid regions. The study proposed two strategies to predict GWL, included an ECa-based GWL model and a remote sensing-based GWL model (RS_GWL), and then compared and explored their performances and cooperation possibilities. To this end, this study first constructed the ECa prediction model and the RS_GWL with ensemble machine learning algorithms using environmental variables and field observations (474 ECa reads and 436 groundwater level observations from a mountain–oasis–desert system, respectively). Subsequently, a strategy to improve the prediction accuracy of GWL was proposed by comparing the correlation between GWL observations and the two models. The results showed that the RS_GWL prediction model explains 30% and 90% of the spatial variability in the two value domain intervals, GWL < 10 m and GWL > 10 m, respectively. The R² of the modeling and the validation of the ECa was 79% and 73%, respectively. Careful analysis of the scatter plots between predicted ECa and GWL revealed that when ECa varies between 0–600 mS/m, 600–800 mS/m, 800–1100 mS/m, and >1100 mS/m, the fluctuation ranges of the corresponding GWL correspond to 0–31 m, 0–15 m, 0–10 m, and 0–5 m. Finally, combining the spatial variability of ECa and RS_GWL spatial distribution map, the following optimization strategies were finally established: GWL < 5 m (in natural land with ECa > 1100 mS/m), GWL < 5 m (occupied by farmland from RS_GWL) and GWL > 10 m (from RS_GWL), and 3 < GWL < 10 m (speculated). In conclusion, this study has demonstrated that the integration of EMI technology has significantly improved the precision of forecasting shallow GWL in oasis plain regions, outperforming the outcomes achieved by the use of remote sensing data alone.

Wheat and corn are two major food crops in Xinjiang. However, the spectral similarity between these crop types and the complexity of their spatial distribution has posed significant challenges to accurate crop identification. To this end, the study aimed to improve the accuracy of crop distribution identification in complex environments in three ways. First, by analysing the kNDVI and EVI time series, the optimal identification window was determined to be days 156–176—a period when wheat is in the grain-filling to milk-ripening phase and maize is in the jointing to tillering phase—during which, the strongest spectral differences between the two crops occurs. Second, principal component analysis (PCA) was applied to Sentinel-2 data. The top three principal components were extracted to construct the input dataset, effectively integrating visible and near-infrared band information. This approach suppressed redundancy and noise while replacing traditional RGB datasets. Finally, the Convolutional Block Attention Module (CBAM) was integrated into the U-Net model to enhance feature focusing on key crop areas. An improved Atrous Spatial Pyramid Pooling (ASPP) module based on deep separable convolutions was adopted to reduce the computational load while boosting multi-scale context awareness. The experimental results showed the following: (1) Wheat and corn exhibit obvious phenological differences between the 156th and 176th days of the year, which can be used as the optimal time window for identifying their spatial distributions. (2) The method proposed by this research had the best performance, with its mIoU, mPA, F1-score, and overall accuracy (OA) reaching 83.03%, 91.34%, 90.73%, and 90.91%, respectively. Compared to DeeplabV3+, PSPnet, HRnet, Segformer, and U-Net, the OA improved by 5.97%, 4.55%, 2.03%, 8.99%, and 1.5%, respectively. The recognition accuracy of the PCA dataset improved by approximately 2% compared to the RGB dataset. (3) This strategy still had high accuracy when predicting wheat and corn yields in Qitai County, Xinjiang, and had a certain degree of generalisability. In summary, the improved strategy proposed in this study holds considerable application potential for identifying the spatial distribution of wheat and corn in arid regions.

Our understanding of water and salt changes in the context of declining groundwater levels in the Tarim Basin remains limited, largely due to the scarcity of hydrological monitoring stations and field observation data. This study utilizes water and salt monitoring data from 474 apparent electromagnetic induction (ECa, measured by EM38-MK2 device) sites across seven oases, combined with groundwater level observation data from representative areas, to analyze the spatiotemporal changes in ECa within the oases of the Tarim Basin from 2000 to 2022. Specific results are shown below: Numerous algorithmic predictions show the ensemble learning algorithm with the smallest error explained 71% of the ECa spatial variability. The ECa was particularly effective at identifying areas where groundwater extends beyond a depth of 5 m, demonstrating increased efficacy when ECa readings exceed the threshold of 1100 mS/m. Our spatiotemporal analysis spanning the years 2000 to 2022 has revealed a significant decline in ECa values within the artificially irrigated zones of the oasis clusters. In contrast, the transitional ecotone between the desert and the oases in Atux, Aksu, Kuqa, and Luntai have experienced a significant increase in ECa value. The variations observed within the defined Zone B, where ECa values ranged from 800 mS/m to 1100 mS/m, and Zone A, characterized by ECa values exceeding 1100 mS/m, aligned with the periodic fluctuations in the groundwater drought index (GDI), indicating a clear pattern of correlation. This study demonstrated that ECa can serve as a valuable tool for revealing the spatial and temporal variations of water resources in arid zones. The results obtained through this approach provided essential references for the local scientific management of soil and water resources.

High-dimensional environmental covariates are increasingly available for digital soil mapping (DSM), but their effective use depends on both the feature-selection strategy and the predictive model architecture. However, systematic evidence remains limited regarding how different metaheuristic feature-selection methods interact with standalone and hybrid learning models across multiple soil and groundwater prediction tasks. This study systematically evaluated the interactions between 10 metaheuristic feature-selection algorithms and 13 predictive models, including random forest (RF), convolutional neural network (CNN), recurrent architectures, CNN–recurrent neural network (RNN) hybrids, squeeze-and-excitation (SE)-enhanced hybrids, and iTransformer-based hybrids, across four prediction tasks involving soil organic carbon (SOC), soil–water extract electrical conductivity (ECe), apparent electrical conductivity (ECa), and groundwater level (GWL) in Xinjiang, China. A total of 149 candidate environmental covariates were considered for ECe, SOC, and ECa, whereas 122 candidate covariates were considered for GWL. The results showed that no single feature-selection method consistently performed best across all four targets; instead, predictive performance depended on the interaction among the feature-selection strategy, predictive architecture, and target variable. CNN–RNN hybrid architectures generally achieved higher predictive performance than standalone models, although their benefits varied among prediction targets. The best-performing combinations yielded coefficient of determination (R²) values of 0.9826, 0.6981, 0.8429, and 0.8085 for GWL, SOC, ECe, and ECa, respectively. These findings indicate that target-specific compatibility, rather than aggressive dimensionality reduction or a universally superior algorithm, is a key determinant of predictive performance in high-dimensional DSM. By demonstrating that feature-selection effectiveness is jointly influenced by model architecture and target characteristics, this study provides a methodological reference for developing target-specific digital soil mapping models in arid regions.