Abstract: Cultivated land can play a vital role in food production and a sustainable ecosystem, in order to maintain the ecological balance and food security. However, numerous cultivated lands have been abandoned in recent years, due to rapid urbanization and industrialization. Efficient and accurate identification of abandoned farmland is crucial to maintaining food security and efficient land use, particularly in the region with high population density and limited agricultural space. This study aims to extract the abandoned farmland from the remote sensing images using SHAP (Shapley Additive exPlanations)-interpretable feature optimization. The study area was taken as Ping’an Town in Fengjie County, Chongqing, China. The sloping farmland was concentrated and susceptible to abandonment, due to the challenging planting conditions on steep slopes. The data source consisted of 2-meter GF satellite images covering the study area, with the imaging time in 2019 and 2020. Then, 34 features were obtained after preprocessing, including the spectral, spectral index, texture, neighborhood, and topographic features. Spearman correlation coefficients and SHAP were also combined to optimize and then remove the redundant features. Six combinations of features were classified and evaluated after optimization. The best combination was selected for the classification of land cover. Finally, a spatial distribution pattern of abandoned farmland was generated after classification and comparison. The result indicated: 1) The topographic, texture, and neighborhood features were effectively improved with the high accuracy of classification, according to the spectral and spectral index features. The topographic features contributed the most. The most representative features were effectively optimized for the classification tasks. Thereby, the information redundancy and overfitting were avoided to improve the accuracy and efficiency of the model. Spearman correlation coefficients were used to filter out the high-correlation features. While the SHAP provided the high-precise feature importance analysis. The removal of weak classification features was realized after combination. The overfitting was also reduced to ensure that only the most representative features were used to improve the computational efficiency and stability. Overall, the features were selected mainly as the slope, red band, green band, blue band, and difference vegetation index. 2) Among the 6 schemes, all algorithms performed best under Scheme 6 (Spearman + SHAP), with an average overall accuracy (OA) of 90.24% in 2019 and 95.83% in 2020. The multiple classifications greatly contributed to the accuracy. The classification was compared with the optimal combination of Scheme 6. Ultimately, the CatBoost achieved the highest accuracy, with an OA of 92.14% and a Kappa of 87.37% in 2019, and an OA of 97.86% and Kappa of 96.59% in 2020. This superior performance of CatBoost was attributed to the categorical feature encoding, ordered boosting mechanism, and symmetric tree structure. The information loss and data leakage were reduced the overfitting for better adaptability to the complex datasets. 3) The optimal classification was overlaid to determine the abandoned farmland area of 3.55km2, with an abandonment rate of 16.94%. Visual interpretation and field surveys showed that the location accuracy of abandonment farmland was achieved at 92.71% after accuracy verification, while the area overlap rate was 83.76%. The smaller patches were measured between 400 and 800 m², where the overlap rate was slightly lower at 82.68%. The abandoned farmland was extracted with high accuracy and low omission rates. The abandoned farmland was effectively fragmented and then verified the feasibility of the optimization. In summary, high accuracy and applicability were obtained to extract the abandoned farmland. The finding can offer valuable technical support and reference to monitor the abandoned farmland, according to the land cover extraction. This approach can also enhance understanding of the decision-making on land use in modern agriculture.