Integrating phenology knowledge and Bayesian networks to map cropping patterns in South China

Crop maps play a crucial role in agricultural remote sensing applications at both regional and national levels, particularly in monitoring cropland use, simulating cropping intensity, estimating crop yields, and assessing agricultural sustainability. Existing crop mapping methods primarily rely on machine learning algorithms, which often depend heavily on sample data and lack portability. Crops follow a relatively stable seasonal growth pattern, which can be captured through time-series remote sensing data. The similarity in phenological characteristics of crops with the same cropping pattern and the differences between those of different cropping patterns serve as the foundation for crop mapping. This research proposes a new method for crop mapping by integrating phenological knowledge into a Bayesian network framework. By extracting key phenological features and encoding crop phenology knowledge with a small number of samples, a Bayesian network model was developed for mapping cropping patterns on the Jianghan Plain. Several spectral and geophysical metrics from key phenological stages were used as feature nodes. The method was validated on the Jianghan Plain, which has a complex cropping pattern with diverse crop types, including winter wheat, winter rapeseed, paddy rice, soybean, and corn. The method demonstrates excellent performance, achieving an overall accuracy of 93 % and a Kappa coefficient of 0.92. The results demonstrate that: (1) Phenological knowledge allows model parameters to be derived without the need for samples (or using very few samples), with no significant reduction in accuracy. (2) The method exhibits strong robustness and portability. The proposed approach enables "weak learning and strong inference," eliminating inaccurate fitting during the inference process and enhancing both the interpretability and portability of the model.