In order to explore the spatiotemporal changes and driving factors of soil organic carbon (SOC) in the agro-pastoral ecotone of northern China, we took Aohan banner, Chifeng City, Inner Mongolia Autonomous Region as the study area, used the random forest (RF) method to predict the SOC from 1989 to 2018, and the geographic detector method (GDM) was applied to analyze quantitatively the natural and anthropogenic factors that are affecting Aohan banner. The results indicated that: (1) After adding the terrain factors, the R² and residual predictive deviation (RPD) of the RF model increased by 1.178 and 0.39%, with root mean square errors (RMSEs) of 1.42 g/kg and 1.05 g/kg, respectively; (2) The spatial distribution of SOC was higher in the south and lower in the north; the negative growth of SOC accounted for 55.923% of the total area, showing a trend of degradation; (3) Precipitation was the main driving factor of SOC spatial variation in the typical agro-pastoral ecotone of northern China, which was also affected by temperature, elevation, soil type and soil texture (p < 0.01). (4). Anthropogenic factors (carbon input and gross domestic product (GDP)) had a greater impact on SOC than did climate factors (temperature and precipitation), making anthropogenic factors the dominant factors affecting SOC temporal variation (p < 0.01). The results of this work constitute a basis for a regional assessment of the temporal evolution of organic carbon in the soil surface, which is a key tool for monitoring the sustainable development of agropastoral ecotones.

Soil organic matter (SOM) is very important to the quality evaluation of cultivated land, especially in fertile black soil areas. Many studies use remote sensing images combined with different machine learning algorithms to predict the regional SOM content. However, the information provided by remote sensing images in different time windows is very different. Taking Youyi Farm, a typical black soil area in Northeast China, as the research area, this study obtains all available Sentinel-2 images covering the research area from 2019 to 2021, calculates the spectral index of single-phase and multi-temporal synthesis images, takes the spectral index and band of each image as the input, and employs the random forest regression algorithm to evaluate the performance of SOM prediction using remote sensing images with different time windows. The results show that: (1) the accuracy of SOM prediction using image band and spectral index is generally improved compared to using only the band; (2) when using single-phase images, the R² range of SOM prediction using image band and spectral index is from 0.16 to 0.59 and the RMSE ranges from 0.82% to 1.23%; When using multi-temporal synthesis images, the R² range of SOM prediction using image band and spectral index is from 0.18 to 0.56 and the RMSE ranges from 0.85% to 1.19%; (3) the highest accuracy of SOM prediction using synthetic images is lower than that of single-phase images; (4) the best time window of the bare soil period in the study area is May. This study emphasizes the importance of the time window to SOM prediction. In subsequent SOM prediction research, remote sensing images with appropriate time windows should be selected first, and then the model should be optimized.

Soil organic matter (SOM) is an essential component of soil and is crucial for increasing agricultural production and soil fertility. The combination of hyperspectral remote sensing and deep learning can be used to predict the SOM content efficiently, rapidly, and cost-effectively on various scales. However, determining the optimal groups, inputs, and models for reducing the spatial heterogeneity of soil nutrients in large regions and to improve the accuracy of SOM prediction remains a challenge. Hyperspectral reflectance data from 1477 surface soil samples in Northeast China were utilized to evaluate three grouping methods (no groups (NG), traditional grouping (TG), and spectral grouping (SG)) and four inputs (raw reflectance (RR), continuum removal (CR), fractional-order differentiation (FOD), and spectral characteristic parameters (SCPs)). The SOM prediction accuracies of random forest (RF), convolutional neural network (CNN), and long short-term memory (LSTM) models were assessed. The results were as follows: (1) The highest accuracy was achieved using SG, SCPs, and the LSTM model, with a coefficient of determination (R2) of 0.82 and a root mean squared error (RMSE) of 0.69%. (2) The LSTM model exhibited the highest accuracy in SOM prediction (R2 = 0.82, RMSE = 0.89%), followed by the CNN model (R2 = 0.72, RMSE = 0.85%) and the RF model (R2 = 0.69, RMSE = 0.91%). (3) The SG provided higher SOM prediction accuracy than TG and NG. (4) The SCP-based prediction results were significantly better than those of the other inputs. The R2 of the SCP-based model was 0.27 higher and the RMSE was 0.40% lower than that of the RR-based model with NG. In addition, the LSTM model had higher prediction errors at low (0–2%) and high (8–10%) SOM contents, whereas the error was minimal at intermediate SOM contents (2–8%). The study results provide guidance for selecting grouping methods and approaches to improve the prediction accuracy of the SOM content and reduce the spatial heterogeneity of the SOM content in large regions.