Soil pH is a key factor that controls soil nutrient availability, soil microbial activities, and crop growth and development. However, studies on the soil pH variations of cultivated lands in different horizons at the regional scale remain limited. In this work, 348 soil samples were collected from three soil horizons (A, B, and C) at 120 sites over the hilly region of Chongqing, southwestern China. Six topographic indicators, four climate parameters, and parent material were considered. Classification and regression trees (CARTs) were applied to investigate the relationships between soil pH and the variables in the A, B, and C horizons. Model performances were evaluated by root mean square error (RMSE), relative root mean square error (RRMSE), and coefficient of determination (R²). Results showed that soil pH increased obviously from the A to C horizons. Soil pH was predicted well by the forcing factors with the CART models in all horizons. RMSE, RRMSE, and R² varied between 0.37 and 0.435, between 5.93 and 7.23%, and between 0.71 and 0.80, respectively. The relative importance of the studied variables to soil pH differed with the horizons. Annual temperature range (ATR), terrain wetness index (TWI), and Melton ruggedness number were critical factors that controlled soil pH variability in the A horizon. Parent material, precipitation of warmest quarter (PWQ), ATR, and TWI were important variables in the B horizon. Parent material, PWQ, ATR, and precipitation were key factors in the C horizon. The results are expected to provide valuable information for designing appropriate measurements for agricultural practices and preventing soil acidification.

Rice is an important agricultural crop in the Southwest Hilly Area, China, but there has been a lack of efficient and accurate monitoring methods in the region. Recently, convolutional neural networks (CNNs) have obtained considerable achievements in the remote sensing community. However, it has not been widely used in mapping a rice paddy, and most studies lack the comparison of classification effectiveness and efficiency between CNNs and other classic machine learning models and their transferability. This study aims to develop various machine learning classification models with remote sensing data for comparing the local accuracy of classifiers and evaluating the transferability of pretrained classifiers. Therefore, two types of experiments were designed: local classification experiments and model transferability experiments. These experiments were conducted using cloud-free Sentinel-2 multi-temporal data in Banan District and Zhongxian County, typical hilly areas of Southwestern China. A pure pixel extraction algorithm was designed based on land-use vector data and a Google Earth Online image. Four convolutional neural network (CNN) algorithms (one-dimensional (Conv-1D), two-dimensional (Conv-2D) and three-dimensional (Conv-3D_1 and Conv-3D_2) convolutional neural networks) were developed and compared with four widely used classifiers (random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM) and multilayer perceptron (MLP)). Recall, precision, overall accuracy (OA) and F1 score were applied to evaluate classification accuracy. The results showed that Conv-2D performed best in local classification experiments with OA of 93.14% and F1 score of 0.8552 in Banan District, OA of 92.53% and F1 score of 0.8399 in Zhongxian County. CNN-based models except Conv-1D provided more desirable performance than non-CNN classifiers. Besides, among the non-CNN classifiers, XGBoost received the best result with OA of 89.73% and F1 score of 0.7742 in Banan District, SVM received the best result with OA of 88.57% and F1 score of 0.7538 in Zhongxian County. In model transferability experiments, almost all CNN classifiers had low transferability. RF and XGBoost models have achieved acceptable F1 scores for transfer (RF = 0.6673 and 0.6469, XGBoost = 0.7171 and 0.6709, respectively).

Soil organic carbon (SOC) is a key property for evaluating soil quality. SOC is thus an important parameter of agricultural soils and needs to be regularly monitored. The aim of this study is to explore the potential of synthetic aperture radar (SAR) satellite imagery (Sentinel-1), optical satellite imagery (Sentinel-2), and digital elevation model (DEM) data to estimate the SOC content under different land use types. The extreme gradient boosting (XGboost) algorithm was used to predict the SOC content and evaluate the importance of feature variables under different land use types. For this purpose, 290 topsoil samples were collected and 49 features were derived from remote sensing images and DEM. Feature selection was carried out to prevent data redundancy. Coefficient of determination (R²), mean absolute error (MAE), mean squared error (MSE), percent root mean squared error (%RMSE), ratio of performance to interquartile range (RPIQ), and corrected akaike information criterion (AICc) were employed for evaluating model performance. The results showed that Sentinel-1 and Sentinel-2 data were both important for the prediction of SOC and the prediction accuracy of the model differed with land use types. Among them, the prediction accuracy of this model is the best for orchard (R² = 0.86 and MSE = 0.004%), good for dry land (R² = 0.74 and MSE = 0.008%) and paddy field (R² = 0.66 and MSE = 0.009%). The prediction model of SOC content is effective and can provide support for the application of remote sensing data to soil property monitoring.

Understanding the spatial variability of soil texture classes is essential for agricultural management. Sentinel-2 data offer valuable vegetation information as proxies for soil properties inference. However, the applications of them in soil texture classification are still limited. This study investigated the usefulness of Sentinel-2 data for predicting soil texture class using an interpretable machine learning (ML) strategy. Multi-temporal Sentinel-2 images were used to get exhaustive vegetation cover information. Basic digital elevation map (DEM) derivatives and stratum were extracted. Three support vector machines with different input parameters (purely DEM derivatives and stratum, purely Sentinel-2, and Sentinel-2 plus DEM derivatives and stratum) were developed. Moreover, in order to improve the transparency in black box ML models, the novel SHapley Additive exPlanations (SHAP) method was applied to interpret the outputs and analyze the importance of individual variables. Results showed that the model with all variables provided desirable performance with overall accuracy (OA) of 0.8435, F1-score of 0.835, kappa statistic of 0.7642, precision of 0.8388, recall of 0.8355 and area under the curve (AUC) of 0.9451. The model with purely Sentinel-2 data performed much better than that with solely DEM derivatives and stratum. The contributions of Sentinel-2 data to explain soil texture class variability were about 17, 41 and 28% for sandy, loamy and clayey soils, respectively. The SHAP method visualized the decision process of ML and indicated that elevation, stratum, and red-edge factors were critical variables. This study offered much-needed insights for the applications of Sentinel-2 in digital soil mapping and ML-assisted tasks.