An NSGA-II-XGBoost Machine Learning Approach for High-Precision Cropland Identification in Highland Areas: A Case Study of Xundian County, Yunnan, China

Abstract

Highlights: What are the main findings? By integrating the NSGA-II optimization framework with the XGBoost algorithm, the proposed model markedly enhances both the accuracy and generalization capability of cropland recognition. It performs exceptionally well in distinguishing croplands from other land types—especially those with similar spectral characteristics or ambiguous boundaries—in plateau areas. What are the implications of the main findings? This study effectively boosts cropland classification accuracy in high-altitude and complex terrain regions through the integration of spectral, radar, and topographic features (e.g., slope and elevation). When combined with the percentage-based method and texture features applied in Google Earth Engine (GEE), this data fusion strategy further confirms the critical role of topographic factors and other auxiliary features in high-precision cropland identification. Accurate identification of cultivated land in plateau and mountainous regions remains challenging due to complex terrain and the fragmented, small-scale distribution of farmland. This study develops a high-precision cropland identification model tailored to such environments, aiming to advance precision agriculture and support the scientific planning and refined management of agricultural resources. Taking Xundian County, Yunnan Province, as a case study, multispectral, synthetic aperture radar (SAR), topographic, texture, and time-series features were integrated to construct a comprehensive multi-source feature space. A baseline land use map was generated by fusing datasets from the European Space Agency (ESA), the Environmental Systems Research Institute (ESRI), and the China Resource and Environment Data Cloud (CRLC). Using 4000 randomly selected sample points, five machine learning algorithms—Support Vector Machine (SVM), Random Forest (RF), Tabular Multiple Prediction (TABM), XGBoost, and the NSGA-II optimized XGBoost (NSGA-II-XGBoost)—were compared for cropland identification. Results show that the NSGA-II-XGBoost model consistently achieved superior performance in classification accuracy, stability, and adaptability, reaching an overall accuracy of 95.75%, a Kappa coefficient of 0.91, a recall of 0.96, and an F1-score of 0.96. These findings demonstrate the strong capability of the NSGA-II-XGBoost model for cropland mapping under complex topographic conditions, providing a robust technical framework and methodological reference for farmland protection and natural resource classification in other mountainous regions.