Quantitative analysis of the spatial non-stationary characteristics of soil salinization influencing factors and the prediction of its spatial distribution are of great significance for the rational use of coastal saline soil resources and the formulation of local prevention and control measures. In this study, the Hekou District of Dongying City, Shandong Province, was used as the study area, and the descriptive statistics of soil salinization status were conducted using classical statistical methods. Spatial autocorrelation theory was used to explore the characteristics of global and local spatial structure of soil salinization in the study area. Influential factors related to soil salinity were selected, and multivariate linear regression （MLR）, geographically weighted regression （GWR）, and multi-scale geographically weighted regression （MGWR） methods were used to model and predict the spatial distribution of soil salinity in the study area and to analyze the spatial heterogeneity of the effects of different influencing factors on soil salinity. The results showed that： ① The mean value of soil salinity in the study area was 5.84 g·kg-1, indicating severe salinization, with a global Moran\'s I index of 0.19 （P<0.00） and obvious spatial aggregation characteristics. ② Among the three models, the MGWR model had the highest modeling accuracy. Compared with that of the MLR model, the Radj2 of GWR and MGWR improved by 0.05 and 0.07, respectively, and the RSS decreased by 210.13 and 179.95, respectively. ③ The results of MGWR regression showed that the spatial distribution of soil salinity appeared to be mainly affected by the middle soil salinity, soil clay content, and vegetation cover from the mean values of standardized regression coefficients of different influencing factors. Different influencing factors had significant spatial non-stationary characteristics on soil salinization. ④ The results of the spatial distribution prediction of soil salinity in MGWR showed that the areas of high soil salinity （≥6 g·kg-1） were mainly distributed in the northern part of the study area, with an overall spatial trend of decreasing from the coast to the interior. The results of the study can be used as a reference for the analysis and predictive mapping of factors affecting soil salinization in the county and on a larger scale using MGWR.

Soil organic carbon (SOC) plays a vital role in global carbon cycling and sequestration, underpinning the need for a comprehensive understanding of its distribution and controls. This study explores the importance of various covariates on SOC spatial distribution at both local (up to 1.25 km) and continental (USA) scales using a deep learning approach. Our findings highlight the significant role of terrain attributes in predicting SOC concentration distribution with terrain, contributing approximately one-third of the overall prediction at the local scale. At the continental scale, climate is only 1.2 times more important than terrain in predicting SOC distribution, whereas at the local scale, the structural pattern of terrain is 14 and 2 times more important than climate and vegetation, respectively. We underscore that terrain attributes, while being integral to the SOC distribution at all scales, are stronger predictors at the local scale with explicit spatial arrangement information. While this observational study does not assess causal mechanisms, our analysis nonetheless presents a nuanced perspective about SOC spatial distribution, which suggests disparate predictors of SOC at local and continental scales. The insights gained from this study have implications for improved SOC mapping, decision support tools, and land management strategies, aiding in the development of effective carbon sequestration initiatives and enhancing climate mitigation efforts.

Sustainable soil resource management depends on reliable soil information, often derived from \'legacy soil data\' or a combination of old and new soil data. However, the task of harmonizing soil data collected at different times remains a largely unexplored in the literature. Addressing this challenge requires incorporating the temporal dimension into mathematical and statistical models for spatio-temporal soil studies. This study aimed to create a comprehensive framework for harmonizing soil data across various time. We assessed the integration of historical and recent soil data, ranging from 4 to 48 years old data, using soil data recency analysis. To achieve this, we introduced an \'age of data\' attribute, calculating the time difference between soil survey years and the present (e.g., 2022). We applied three machine learning models - Decision Trees (DT), Random Forest (RF), Gradient Boosting (GBM) - to a dataset containing 6339 sites and 28,149 depth-harmonized layers. The results consistently demonstrated robust performance across models, RF outperforming with an R-squared value of 0.99, RMSE of 1.41, and a concordance of 0.97. Similarly, DT and GBM also showed strong predictive power. Terrain-derived environmental covariates played a more important role than land use and land cover (LULC) change in predicting soil data recency. While LULC change showed soil organic carbon concentration variability across the different depths, it was a less important factor. Anthropogenic factors, such as LULC change and normalized difference vegetation index (NDVI), were not primary determinants of soil data recency. Variations in soil depth had no impact on predicting soil data recency. This study validated that terrain-derived covariates, especially elevation factors, effectively explain the quality of older soil data when predicting current soil attributes using the soil data recency concept. This approach has the potential to enhance real-time estimates, such as carbon budgets, and we emphasize its importance in global earth system models.

Soil salinization and sodification, the primary causes of land degradation and desertification in arid and semi-arid regions, demand effective monitoring for sustainable land management. This study explores the utility of partial least square (PLS) latent variables (LVs) derived from visible and near-infrared (Vis-NIR) spectroscopy, combined with remote sensing (RS) and auxiliary variables, to predict electrical conductivity (EC) and sodium absorption ratio (SAR) in northern Xinjiang, China. Using 90 soil samples from the Karamay district, machine learning models (Random Forest, Support Vector Regression, Cubist) were tested in four scenarios. Modeling results showed that RS and Land use alone were unreliable predictors, but the addition of topographic attributes significantly improved the prediction accuracy for both EC and SAR. The incorporation of PLS LVs derived from Vis-NIR spectroscopy led to the highest performance by the Random Forest model for EC (CCC = 0.83, R2 = 0.80, nRMSE = 0.48, RPD = 2.12) and SAR (CCC = 0.78, R2 = 0.74, nRMSE = 0.58, RPD = 2.25). The variable importance analysis identified PLS LVs, certain topographic attributes (e.g., valley depth, elevation, channel network base level, diffuse insolation), and specific RS data (i.e., polarization index of VV + VH) as the most influential predictors in the study area. This study affirms the efficiency of Vis-NIR data for digital soil mapping, offering a cost-effective solution. In conclusion, the integration of proximal soil sensing techniques and highly relevant topographic attributes with the RF model has the potential to yield a reliable spatial model for mapping soil EC and SAR. This integrated approach allows for the delineation of hazardous zones, which in turn enables the consideration of best management practices and contributes to the reduction of the risk of degradation in salt-affected and sodicity-affected soils.

Artificial neural networks (ANNs) have proven to be a useful tool for complex questions that involve large amounts of data. Our use case of predicting soil maps with ANNs is in high demand by government agencies, construction companies, or farmers, given cost and time intensive field work. However, there are two main challenges when applying ANNs. In their most common form, deep learning algorithms do not provide interpretable predictive uncertainty. This means that properties of an ANN such as the certainty and plausibility of the predicted variables, rely on the interpretation by experts rather than being quantified by evaluation metrics validating the ANNs. Further, these algorithms have shown a high confidence in their predictions in areas geographically distant from the training area or areas sparsely covered by training data. To tackle these challenges, we use the Bayesian deep learning approach \"last-layer Laplace approximation\", which is specifically designed to quantify uncertainty into deep networks, in our explorative study on soil classification. It corrects the overconfident areas without reducing the accuracy of the predictions, giving us a more realistic uncertainty expression of the model\'s prediction. In our study area in southern Germany, we subdivide the soils into soil regions and as a test case we explicitly exclude two soil regions in the training area but include these regions in the prediction. Our results emphasize the need for uncertainty measurement to obtain more reliable and interpretable results of ANNs, especially for regions far away from the training area. Moreover, the knowledge gained from this research addresses the problem of overconfidence of ANNs and provides valuable information on the predictability of soil types and the identification of knowledge gaps. By analyzing regions where the model has limited data support and, consequently, high uncertainty, stakeholders can recognize the areas that require more data collection efforts.

Soil properties influence plant physiology and growth, playing a fundamental role in shaping species niches in temperate forest ecosystems. Here, we investigated the impact of soil data quality on the performance of species distribution models (SDMs) of 41 woody plant species in Swiss forests. We compared models based on measured soil properties with those based on digitally mapped soil properties on regional (Swiss Forest Soil Maps) and global scales (SoilGrids). We first calibrated topo-climatic SDMs with measured soil data and plant species presences and absences from mature temperate forest stand plots. We developed further models using the same soil predictors, but with values extracted from digital soil maps at the nearest neighbouring plots of the Swiss National Forestry Inventory. The predictive power of SDMs without soil information compared to those with soil information, as well as measured soil information vs digitally mapped, was evaluated with metrics of model performance and variable contribution. On average, models with measured and digitally mapped soil properties performed significantly better than those without soil information. SDMs based on measured and Swiss Forest Soil Maps showed higher performance, especially for species with an \'extreme\' niche position (e.g., preference for high or low pH), compared to those using SoilGrids. Nevertheless, if no regional soil maps are available, SoilGrids should be tested for their potential to improve SDMs. Moreover, among the tested soil predictors, pH, and clay content of the topsoil layers most improved the predictive power of SDMs for forest woody plants. In conclusion, we demonstrate the value of regional soil maps for predicting the distribution of woody species across strong environmental gradients in temperate forests. The improved accuracy of SDMs and insights into drivers of distribution may support forest managers in strategies supporting e.g. biodiversity conservation, or climate adaptation planning.

Phosphorus (P) is a critical nutrient for primary production in terrestrial and aquatic ecosystems. As P mineral reserves are finite and non-renewable, there is an increasing discussion on its sustainable utilization to safeguard food security for future generations. Understanding the spatial distribution of soil P is central in advancing effective phosphorus management and fostering sustainable agricultural practices. This study aims to digitally map the stocks of available P (AP) and total P (TP) in Brazil at a fine resolution (30 m). Using the Random Forest machine learning algorithm and a database of topsoil (0-20 cm) with 28,572 samples for AP and 3154 for TP, we predicted P stocks based on environmental covariates related to soil formation processes. By dividing Brazil into two sub-regions, representing areas with native coverage and anthropogenic ones, we built independent predictive models for each sub-region. Our results show that Brazil has a TP stock of 531 Tg and an AP stock of 17.4 Tg. The largest soil TP stocks are in the Atlantic Forest biome (73.8 g.m2), likely due to higher organic carbon stocks in this biome. The largest AP stocks were in the Caatinga biome (2.51 g.m2) because of younger soils with low P adsorption capacity. We also found that fertilizer use significantly increased AP stocks in agricultural areas compared to native ones. Our results indicated that AP stocks strongly influenced Brazil\'s agricultural production, with a correlation coefficient ranging from 0.20 for coffee crops to 0.46 for soybean. The maps generated in this study are expected to contribute to the sustainable use of P in agriculture and environmental systems.

UNASSIGNED: Unlike most of Europe, Andalucía in southern Spain as a Mediterranean area still lacks digital maps of SOC content provided by machine learning algorithms. The wide diversity of climate, geology, hydrology, landscape, topography, vegetation, and micro-relief data as easy-to-obtain covariates facilitated the development of digital soil mapping (DSM). The purpose of this research is to model and map the spatial distribution of SOC at three depths, in an area of approximately 10000 km 2 located in Seville and Cordoba Provinces, and to use R programming to compare two machine learning techniques (cubist and random forest) for developing SOC maps at multiple depths.
UNASSIGNED: Environmental covariates used in this research include nine derivatives from digital elevation models (DEM), three climatic variables and finally eighteen remotely-sensed spectral data (band ratios calculated by the acquired Landsat-8 OLI and Sentinel-2A MSI in July 2019). In total, 300 soil samples from 100 points were taken (0-25 cm). The purpose of this research is to model and map the spatial distribution of SOC, in an area with approximately 10000 km2 located in Seville and Cordoba Provinces, and to compare two machine learning techniques (cubist and random forest) by R programming.
UNASSIGNED: The findings showed that the novel approach for integrating the indices using Landsat-8 OLI and Sentinel-2A MSI satellite data had a better result.
UNASSIGNED: Finally, we obtained evidence that the resolution of satellite images is more important in modelling and digital mapping.

Soil organic matter is an important indicator of soil fertility, and it is necessary to improve the accuracy of regional organic matter spatial distribution prediction. In this study, we analyzed the organic matter content of 1 690 soil surface layers (0-20 cm) and collected data on the natural environment and human activities in the Weining Plain of the Yellow River Basin. The SOM spatial distribution prediction model was established with 1 348 points using classical statistics, deterministic interpolation, geostatistical interpolation, and machine learning, respectively, and 342 sample points data were used as the test set to test and analyze the prediction accuracy of different models. The results showed that the average SOM content of the surface soil of the Weining Plain was 14.34 g·kg-1, and the average soil organic matter variation across 1 690 sampling points was 34.81%, indicating a medium degree of variability. The results also revealed a spatial distribution trend, with low soil organic matter content in the northeast and southwest, high soil organic matter on the left and right banks of the Yellow River in the middle, and relatively high soil organic matter in the sloping terrain of the Weining Plain. The four types of methods in order of high to low prediction accuracy were the machine learning method, geostatistical interpolation method, deterministic interpolation method, and classical statistical method. Through comparison, the BP neural network that was improved based on the optimized sparrow search algorithm had the best prediction accuracy, and the optimized sparrow search algorithm had better convergence accuracy, avoided falling into local optimization, prevented data overfitting, and had better prediction ability. This optimization algorithm can improve the accuracy of SOM prediction and has good application prospects in soil attribute prediction.

Assessment and proper management of sites contaminated with heavy metals require precise information on the spatial distribution of these metals. This study aimed to predict and map the distribution of Cd, Cu, Ni, Pb, and Zn across the conterminous USA using point observations, environmental variables, and Histogram-based Gradient Boosting (HGB) modeling. Over 9180 surficial soil observations from the Soil Geochemistry Spatial Database (SGSD) (n = 1150), the Geochemical and Mineralogical Survey of Soils (GMSS) (n = 4857), and the Holmgren Dataset (HD) (n = 3400), and 28 covariates (100 m × 100 m grid) representing climate, topography, vegetation, soils, and anthropic activity were compiled. Model performance was evaluated on 20 % of the data not used in calibration using the coefficient of determination (R2), concordance correlation coefficient (ρc), and root mean square error (RMSE) indices. Uncertainty of predictions was calculated as the difference between the estimated 95 and 5 % quantiles provided by HGB. The model explained up to 50 % of the variance in the data with RMSE ranging between 0.16 (mg kg-1) for Cu and 23.4 (mg kg-1) for Zn, respectively. Likewise, ρc ranged between 0.55 (Cu) and 0.68 (Zn), respectively, and Zn had the highest R2 (0.50) among all predictions. We observed high Pb concentrations near urban areas. Peak concentrations of all studied metals were found in the Lower Mississippi River Valley. Cu, Ni, and Zn concentrations were higher on the West Coast; Cd concentrations were higher in the central USA. Clay, pH, potential evapotranspiration, temperature, and precipitation were among the model\'s top five important covariates for spatial predictions of heavy metals. The combined use of point observations and environmental covariates coupled with machine learning provided a reliable prediction of heavy metals distribution in the soils of the conterminous USA. The updated maps could support environmental assessments, monitoring, and decision-making with this methodology applicable to other soil databases, worldwide.