Knowledge-guided adaptive spatial-temporal graph contrastive learning framework: Regional crop diseases prediction based on electronic medical records.

Abstract

The occurrence of crop diseases exhibits nonlinear and dynamic spatial-temporal correlations. How to realize real-time and accurate regional disease prediction is an emerging challenge in smart agriculture. Existing research is hindered by difficulties in data acquisition, high implementation costs of IoT technologies, the complex nature of disease dynamics, and diverse transmission pathways. Each prescription in Plant Electronic Medical Records (PEMRs) includes data on crops, diseases, environmental factors, and their interrelationships, offering a novel approach for regional disease prediction through big data analysis. In this paper, we quantify regional disease occurrence by counting the number of prescriptions generated in a specific area on a given day, representing disease occurrence networks as graphs and transforming regional disease prediction into a spatial-temporal forecasting problem. We propose a knowledge-guided adaptive spatial-temporal graph contrastive learning framework (called KAST-Graph) for regional disease prediction. By dynamically generating graph structures that integrate adaptive and geographically-informed adjacency matrices, KAST-Graph uncovers explicit and implicit relationships between nodes, enabling the learning of global spatial correlations. This work is the first attempt to incorporate prior knowledge into regional crop disease prediction, enhancing the model's capacity to capture fine-grained information. Additionally, we design a contrastive learning-based enhancement scheme to improve robustness against noise data, missing values, and uneven spatial-temporal distributions. Experimental results show that KAST-Graph outperforms state-of-the-art baselines in spatial-temporal mining on PEMRs big data, achieving the best MAE, RMSE, and MAPE scores of 5.71, 9.50, and 4.56 %. This work holds significant implications for the early warning and prevention of regional crop diseases. The core code of this work can be found from: https://github.com/xc9873/KAST-Graph.git.