Dynamic monitoring and classification identification of crop rotation patterns based on continuous change detection and classification algorithms

Non-grain cultivation on arable land has posed an ever increasing challenge on national food security in sustainable agriculture. Particularly, the winter wheat–summer maize rotation systems dominate cereal production in the North China Plain. It is therefore required to accurately identify and constantly monitor the crop rotation patterns for the land-use transitions. However, existing remote sensing approaches remain dependent largely on single-temporal imagery during optimal phenological windows. Cloud contamination can frequently cause to capture the continuous and nonlinear dynamics of multi-season cropping systems. In this study, a time-series framework was developed to identify the crop rotation patterns from dynamic monitoring data. Continuous Change Detection and Classification (CCDC) algorithm was integrated with machine learning models using dense Sentinel-2 observations. A case study was also selected as the Hua County, Henan Province, China. The representative wheat–maize double-cropping region was characterized by spatial heterogeneity. All available Sentinel-2 Level-2A images were acquired from 2018 to 2024, and then processed on the Google Earth Engine (GEE) platform. The continuous multi-year spectral time series were constructed after image processing. A feature space was designed to characterize the crop phenology and surface conditions. Multi-spectral reflectance bands were incorporated with the six vegetation 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). A systematic comparison was made on two classifications. One was an improved CCDC algorithm. The third-order harmonic regression was fitted into the pixel-level time series to explicitly capture intra-annual growth rhythms and long-term trends. Another was a conventional single-phase approach. Median composites were generated for the key phenological periods of wheat and maize. Annual rotation patterns were then inferred after seasonal overlay. Subsequently, the regression coefficients, harmonic components, amplitudes, and phase parameters from the CCDC model were input for Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) classifiers. The performance was evaluated using five-fold cross-validation and multiple accuracy metrics. The results demonstrate that the CCDC framework was achieved in the high accuracy, robustness and consistency to classify annual rotation patterns in the continuous crop transitions under variable atmospheric conditions. Among them, the highest performance was found in the CCDC–ANN combination, with an average overall accuracy of 91.8% and a Kappa coefficient of 0.891, which was improved by approximately 20% over the conventional approach. The superior performance of the ANN model was also obtained to learn complex nonlinear relationships in dense time-series features. Spatiotemporal analysis further revealed the substantial heterogeneity in the crop rotation patterns. Staple–non-staple rotations were concentrated in western hilly areas with the complex terrain, whereas the stable staple–staple and non-staple–staple systems were dominated in the eastern plains. Temporally, all major rotation types exhibited an “increase–decline–recovery” trajectory from 2018 to 2024. The great variation was primarily attributed to the policy adjustments, market dynamics, and environmental constraints. Overall, the CCDC–ANN framework can provide an accurate and scalable solution to map the wheat–maize rotation systems. The finding can also offer the strong potential to the regional monitoring of non-grain cropland and land use in modern agriculture.