Remote sensing extraction of high-standard farmland in hill–plain transition zones based on multi-feature fusion

High-standard farmland (HSF) is one of the most important natural sources in modern agriculture. Accurate identification and spatial mapping are essential for the land consolidation and the national food security. However, the recognizing HSF with remote sensing can still remain challenging in the hill–plain transition zones. The fragmented landscapes and heterogeneous land cover can obscure the spectral signals of the agricultural land. Particularly, the conventional classification can rely mainly on the natural attributes, such as the spectral and phenological features. Therefore, the HSF cannot be effectively distinguished from the general cultivated land, which is defined by the crop type of the engineering and management standards, such as the parcel regularity, irrigation, and drainage infrastructure. In this study, a multidimensional feature fusion was developed and then validated to integrate the policy-derived spatial attributes with the conventional remote sensing features, particularly for the high-precision HSF extraction. The study area was taken from a representative hill–plain transition zone in the Zhangfeng Town, Dehong Prefecture, Yunnan Province, China. Multi-source datasets were then constructed, including the preprocessed Sentinel-2 and Gaofen-2 imagery with a 10 m Digital Elevation Model (DEM). A 20-dimensional feature set was classified into four feature types: (1) spectral features from ten Sentinel-2 bands; (2) spectral indices including NDVI, RVI, NDWI, and OSAVI to enhance the vegetation and moisture information; (3) phenological features—Start of Season (SOS), End of Season (EOS), and Length of Season (LOS)—derived from time-series NDVI using a threshold and linear interpolation; and (4) regional features—Patch Density (PD) and Ditch Density (DD) to quantify parcel fragmentation and irrigation infrastructure, together with the slope from the DEM. Five schemes (S1–S5) were combined to evaluate the contribution rate of each feature type. Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) machine learning were employed for the classification, according to the identical training and validation samples from the field surveys and high-resolution image interpretation. The results show that the regional features were significantly enhanced the performance of the classification. Compared with the spectral-only scheme (S1), the scheme incorporating regional features (S4) were improved the overall accuracy by 9.41% for RF and 5.60% for XGBoost, indicating the importance of the policy-related spatial characteristics, such as the parcel regularity and ditch networks. The full-fusion scheme (S5) was achieved in the best performance, with the overall accuracies of 96.15% (Kappa = 0.91) for RF and 96.79% (Kappa = 0.93) for XGBoost. Feature importance analysis indicated that the DD and PD were the most influential predictors, thus contributing more than any single spectral or phenological variable. Spatial pattern analysis revealed that the HSF covered 58.44 km2 in the study area, indicating the pattern of the “basin aggregation and hilly fragmentation.” Specifically, the 84.43% of HSF occurred on the slopes less than 15°, and 83.13% was located within 300 m of the major ditch networks. The topographic suitability and irrigation accessibility were consistent with the national HSF construction. The policy-derived regional features were quantified to integrate with the multi-source remote sensing data and advanced machine learning, such as the XGBoost. The accuracy of the HSF identification was greatly improved in the complex terrains. The framework can provide a reliable technical solution to monitor the HSF construction. The finding can also offer the scientific support for the precision land management in mountainous areas of the sustainable agriculture.