Abstract: Soil erosion regionalization has been one of the most effective strategies in erosion control. Conventional approaches can rely on clustering spatial units according to similar multivariate features at a single point in time. Only a static snapshot of the regional states can be captured in the widely used approaches, such as the K-means, Fuzzy C-Means (FCM), and Self-Organizing Feature Maps (SOFM). The inherent spatiotemporal dynamics cannot consider the evolution of soil erosion, which results from the long-term, coupled interactions of climatic factors (precipitation and temperature), terrain, soil properties, and human activities. Dynamic regional control has been limited in its effectiveness. In this study, a soil erosion regionalization model was introduced using geographical process theory: the Geographical Process Regionalization of Soil Erosion (GPR-SE). The GPR-SE model was used to explicitly capture the spatiotemporal evolution patterns of soil erosion. Both spatial heterogeneity and temporal process similarity were quantified to construct the integrated spatiotemporal features for the dynamic erosion. Static indicators were processed to extract the high-level, compressed spatial features using a Convolutional Autoencoder (CAE) with hierarchical representation learning. Noise and redundancy were effectively eliminated from the raw spatial data to preserve the essential spatial patterns. Concurrently, dynamic indicators – specifically multi-decadal precipitation and temperature time series – were analyzed with a Long Short-Term Memory (LSTM) network. The LSTM was used to capture the complex temporal dependencies and non-linear evolution patterns within sequential climate data. The gating mechanisms and memory cells were used to avoid the information loss inherent in conventional statistical aggregation. The spatial features from the CAE and temporal embeddings from LSTM were fused according to the feature concatenation, indicating the spatiotemporal representations of the geographical erosion. The fused spatiotemporal features were input for the clustering analysis. Three prominent algorithms – K-means, FCM, and SOFM – were evaluated using the Silhouette Coefficient (SiL) and Davies-Bouldin Index (DBI). The optimal model was identified to maximize the inter-region dissimilarity and intra-region homogeneity. Therefore, the K-means served as the superior algorithm when clustering into five regions. The optimal GPR-SE framework was applied to Hailun City, a representative black soil area of erosion in Northeast China. Post-regionalization, a Random Forest (RF) was employed to identify the dominant differentiation drivers and quantify feature importance after permutation contribution analysis. The results indicate: 1) Optimal Clustering: The cluster number was set to 5. The K-means algorithm performed optimally, according to the Sil and DBI metrics. The best balance was achieved to maximize the inter-group differences, while minimizing intra-group differences within the spatiotemporal feature space. The performance outperformed the FCM and SOFM for this specific task and dataset. 2) Dominant Differentiation Factors: RF analysis revealed that the dynamic factors of soil erosion were the primary drivers of the regional differentiation under the GPR-SE framework. Precipitation and temperature also exhibited the highest contribution rates (25.08% and 21.65%, respectively), indicating the critical role of the climatic dynamics in the spatiotemporal patterns of the erosion over time. Landform type (14.05%) and soil erosion modulus (11.95%) were significant secondary factors, while the static soil properties and land use played the lesser roles. 3) Defined Erosion Process Zones: the key process factors (primarily precipitation and temperature dynamics) and erosion intensity were combined to partition the study area. Four types of soil erosion regions were obtained: Low Precipitation-Medium Temperature-Moderate Erosion Zone, Low Precipitation-Low Temperature-Light Erosion Zone, Low Precipitation-High Temperature-Severe Erosion Zone, and Low Precipitation-Low Temperature-Moderate Erosion Zone, indicating the combination of the climatic dynamics and erosion severity. The GPR-SE model shifted the paradigm from the static, attribute-based partitioning to dynamic, process-oriented regionalization. The spatiotemporal evolution of the key erosion drivers was effectively captured, particularly climatic factors. A more accurate delineation of soil erosion zones can be represented by their inherent evolution. This finding can provide a robust approach to preventing soil erosion for the targeted conservation and resource allocation in erosion-prone regions, like the critical black soil areas of Northeast China and beyond.