How can we identify abandoned cropland and secure food safety?

Abandoned cropland poses a threat to food security and sustainable land use. But how can we identify them efficiently? A new study has found an innovative solution. Scientists combined three land cover datasets: ESA WorldCover, ESRI Land Cover, and CRLC. Using Google Earth Engine, they extracted features from these data and created high-quality training samples. To optimize the model, they used a genetic algorithm (NSGA-II) combined with XGBoost, improving identification speed by 75% without losing accuracy. The model achieved an accuracy of 98.17% and a Kappa score of 0.9633. To interpret the results, they used SHAP and LIME. SHAP showed that features like SAVG, B3_p25, Road, DEM, and Population are the most important. LIME revealed that some areas have different local influences, such as slope and SAVG. This study offers a practical framework for identifying and managing abandoned cropland, balancing precision, interpretability, and efficiency.

The proposed model combines NSGA-II and XGBoost to optimize the classification of abandoned cropland. NSGA-II is a genetic algorithm that finds optimal solutions for problems with multiple objectives, while XGBoost is a boosting algorithm that improves model accuracy. The combination of these two methods allowed for reduced computation time and improved accuracy. SHAP and LIME are interpretability techniques that help understand how the model makes decisions. SHAP evaluates the global importance of features, while LIME explains local model decisions. This global-local approach provides a comprehensive view of the factors influencing cropland abandonment.