关键词: 耕地质量评价, 卡方分箱, 合成少数类过采样技术, 集成学习

Abstract: In order to fully understand the county‑level cultivated land resources，improve the efficiency and accuracy of cultivated land quality evaluation，and reduce the influence of human subjective factors，we selected Shanzhou District of Sanmenxia City in Henan Province as the research object and graded its cultivated land quality in this paper.The characteristic variables were divided into categorical variables and numerical variables according to the characteristics of data. For numerical variables，the Chi‑square binning method was used for processing.The processed results and categorical variables were assigned by one‑hot encoding，and the SMOTE method was used to balance the sample data for data with a small number of sample categories.Then the method of machine learning was used，and the quality of arable land was graded using XGBoost，LightGBM，ANN，and a combined XGBoost‑LightGBM‑ANN model.The results showed that，when the processed unsampled data were used as the training set and substituted into the single models as well as combined models for training and prediction of cropland quality evaluation，metrics such as accuracy，precision，recall，and F1 scores all reached values of over 0. 97.Among them，the values of the indicators corresponding to the single models ANN，LightGBM and XGBoost were increasing，indicating that the application of machine learning models in the evaluation of cropland quality produced good results.Compared with the unsampled data，the sampled data as a training set produced a significant improvement in the model performance.In the prediction of single models and the combined model，the values of each index reached more than 0.99，and the XGBoost‑LightGBM‑ANN combined model based on the integrated learning voting algorithm had an obvious advantage，with an accuracy of 0.998 3.

Key words: Cultivated land quality evaluation, Chi?square binning, SMOTE, Ensemble learning