To establish a machine learning predictive model based on multidimensional spatiotemporal risk feature engineering for forecasting the results of supervisory sampling inspections of edible agricultural products.

Using supervisory sampling data for edible agricultural products from Fujian’s prefecture-level cities between 2022 and 2024, information on product type, sampling location, sampling site and production date were systematically integrated. Multidimensional spatiotemporal risk feature engineering was applied to process the dataset, constructing 6 kinds of new features to form a novel dataset based on the original data. The 3 kinds of mainstream binary classification prediction models were applied to forecast inspection outcomes using the processed data.

Across all 3 kinds of mainstream models, the multidimensional spatiotemporal risk feature engineering method demonstrated superior predictive performance compared to traditional sampling methods.Specifically, the combination of multidimensional spatiotemporal risk feature engineering with light gradient boosting machine (LightGBM) achieved the best overall performance on the test set samples, with an AUC of 0.845,Recall of 0.881, G-mean of 0.741 and misclassification costs of 8561.

Compared to traditional sampling methods, multidimensional spatiotemporal risk feature engineering enhances the model’s ability to capture complex risk patterns, leading to a comprehensive improvement in predictive performance. The LightGBM model based on multidimensional spatiotemporal risk feature engineering can accurately predict the inspection results of edible agricultural products, providing technical support for formulating differentiated, dynamic supervision and inspection plans and risk early warning systems for edible agricultural products.