Dynamic Monitoring and Classification Identification of Crop Rotation Patterns Based on Continuous Change Detection and Classification Algorithms

The rapid expansion of non-grain cultivation on arable land poses a growing challenge to food security and sustainable agricultural management, particularly in the North China Plain where winter wheat–summer maize rotation systems dominate cereal production. Reliable identification and long-term monitoring of crop rotation patterns are therefore critical for assessing land-use transitions and supporting evidence-based policy interventions. Existing remote sensing approaches, however, remain largely dependent on single-temporal imagery acquired during optimal phenological windows, which are frequently affected by cloud contamination and are insufficient to capture the continuous and nonlinear dynamics of multi-season cropping systems. To address these limitations, this study develops a time-series–based framework for dynamic crop rotation mapping by integrating the Continuous Change Detection and Classification (CCDC) algorithm with machine learning models using dense Sentinel-2 observations. Hua County, Henan Province, a representative wheat–maize double-cropping region characterized by pronounced spatial heterogeneity, was selected as the case study. All available Sentinel-2 Level-2A images from 2018 to 2024 were processed on the Google Earth Engine (GEE) platform to construct a continuous, multi-year spectral time series. A comprehensive feature space was designed to characterize crop phenology and surface conditions, incorporating multi-spectral reflectance bands and six vegetation-related indices, including the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Bare Soil Index (BSI), Yellow Index, Normalized Difference Red Edge Index (NDREI), and Inverted Red-Edge Chlorophyll Index (IRECI). Two contrasting classification strategies were implemented and systematically compared. The first strategy followed a traditional single-phase approach, generating median composites for key phenological periods of wheat and maize and inferring annual rotation patterns through seasonal overlay. The second strategy employed an improved CCDC algorithm, fitting third-order harmonic regression models to pixel-level time series to explicitly capture intra-annual growth rhythms and long-term trends. Regression coefficients, harmonic components, amplitudes, and phase parameters derived from the CCDC model were subsequently used as inputs for Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) classifiers. Model performance was evaluated using five-fold cross-validation and multiple accuracy metrics. The results reveal pronounced differences between the two strategies. Although the traditional single-phase method achieved high classification accuracy for individual growing seasons (overall accuracy up to 96.8%, Kappa coefficient 0.96), its performance declined substantially when identifying annual rotation patterns, yielding an average overall accuracy of only 71.3%. This degradation is primarily attributed to the inability of single-temporal composites to represent continuous crop transitions under variable atmospheric conditions. In contrast, the CCDC-based framework demonstrated markedly improved robustness and consistency. Among the tested models, the CCDC–ANN combination achieved the highest performance, with an average overall accuracy of 91.8% and a Kappa coefficient of 0.891, representing an improvement of approximately 20% over the traditional approach. The superior performance of the ANN model highlights its capacity to learn complex nonlinear relationships embedded in dense time-series features. Spatiotemporal analysis further revealed substantial heterogeneity in crop rotation patterns. Staple–non-staple rotations were concentrated in western hilly areas with complex terrain, whereas stable staple–staple and non-staple–staple systems dominated the eastern plains. Temporally, all major rotation types exhibited a “increase–decline–recovery” trajectory from 2018 to 2024, reflecting the combined effects of policy adjustments, market dynamics, and environmental constraints. Overall, this study demonstrates that the proposed CCDC–ANN framework provides an accurate, scalable, and automated solution for mapping wheat–maize rotation systems, offering strong potential for regional to national applications in non-grain cropland monitoring and agricultural land-use management.