Mapping of Cropping Patterns on the Jianghan Plain based on Phenological Knowledge and Bayesian Networks

The spatial distribution map of crops is essential for a wide range of agricultural and environmental applications, including but not limited to monitoring land use patterns, simulating cropping intensity, estimating grain yields, and assessing the sustainability of agricultural systems. It provides a foundational geospatial data layer that supports decision-making processes in agricultural planning, policy formulation, and food security evaluation. Current crop mapping techniques primarily rely on machine learning algorithms, which depend heavily on training data and struggle with portability. Deep learning methods, while powerful, also require large amounts of training data, which exacerbates the issue of data dependency. However, the challenge of obtaining reliable, accurate, and sufficiently large labeled datasets has consistently been a key limitation for data-driven methods, particularly in large-scale or heterogeneous agricultural regions. The scarcity of high-quality training data is compounded by cloud contamination in optical remote sensing, inconsistent field surveys, and phenological variability across landscapes, making scalable crop mapping a persistent challenge. Crops follow a relatively stable and biologically driven growth rhythm governed by environmental conditions and agronomic practices. This growth cycle exhibits characteristic phenological stages (e.g., emergence, vegetative growth, flowering, maturity), which can be systematically captured and quantified using time-series remote sensing data, especially vegetation indices such as NDVI or EVI. These temporal dynamics offer a unique opportunity to encode expert phenological knowledge into computational models in a way that reduces dependence on large training data. In this study, we introduce a novel cropping pattern mapping method based on Bayesian Networks, incorporating crop phenology knowledge into the classification framework. By extracting key phenological features corresponding to critical growth stages, and using only a small number of representative training samples, we perform knowledge probabilistic encoding to define conditional dependencies and construct a Bayesian Network tailored for phenology-driven crop type classification. Empirical experiments were conducted in a region with highly complex cropping patterns to validate the effectiveness of the proposed method. The results demonstrate that: 1) With phenological knowledge as a guiding framework, model parameters can be established either without training data or with only a limited number of samples, thus maintaining accuracy under conditions of sample scarcity, with an overall mapping precision exceeding 92%. This demonstrates that prior knowledge can effectively serve as a surrogate for data in knowledge-based remote sensing; 2) the proposed Bayesian Network classification framework exhibits a 'weak learning - strong inference' capability, whereby the model avoids overfitting to limited samples and instead leverages domain-specific knowledge structures for inference. In this manner, the imprecise fitting often observed in data-driven models is eliminated or neutralized during the inference phase, reducing the dependency of the machine learning method on data by 42%. As a result, the method not only mitigates the data dependency and overfitting issues inherent in traditional machine learning models, but also enhances the interpretability, transparency, and portability of the classification process. Its ability to generalize across different temporal and spatial contexts demonstrates strong robustness, especially in regions where training data are sparse or costly to acquire. The integration of domain knowledge with probabilistic graphical modeling represents a promising pathway for crop mapping in data-constrained environments and offers a practical alternative to fully data-driven approaches in remote sensing classification tasks.