Integrating phenology knowledge graph into parcel-scale crop classification using multi-period deep time series modelling

• A classification framework is developed for parcel-scale crop type mapping. • Crop knowledge graph is auto-generated using decision tree mechanism. • Multi-period feature aggregation allows different crop growth pattern recognition. • Knowledge graph-embedded approach improves crop classification accuracy. Parcels are the fundamental units of agricultural management; accurate crop classification of cropland parcels is crucial for the implementation of precision agriculture. Despite extensive knowledge of crop growth processes, the difficulty in acquiring this knowledge and its modal differences with remote sensing data hinder its application in crop classification research. Moreover, the highly complex and variable growth patterns of crops present significant challenges for time-series crop classification. We propose a novel crop classification framework that extracts intricate multi-period features of crop growth from remote sensing time-series signals. Additionally, we introduce an automatic construction process for crop remote sensing knowledge graphs based on a decision tree structure, capturing the association between crops and remote sensing time-series data. Through graph convolution, knowledge graph serves as a global guide to improve crop classification. By combining field survey samples with visible, near-infrared, and radar signals, we constructed a parcel-scale dataset of rice and wheat crops across four cities in the middle and lower reaches of the Yangtze River using zonal feature aggregation methods for evaluation. The results indicate that the proposed framework achieves accuracies ranging from 89.45 % to 94.43 % across the four datasets. We conducted inferences in the four cities and compared the results with county-level statistical data, achieving R 2 values of 0.89 and 0.97 for wheat and rice planting areas, respectively. Our proposed framework can automatically generate crop knowledge graphs based on samples from different regions, overcoming the modal barriers between the knowledge space and the remote sensing feature space, thus enhancing crop recognition accuracy.