基于无人机多源数据与机器学习协同的冻胀丘高精度识别及多尺度空间分布格局OA

Frost mounds,a prevalent periglacial landform in permafrost regions,exhibit significant spatiotemporal heterogeneity in their developmental processes in response to accelerating climate warming trends.Such alterations potentially exert profound impacts on the structural and functional integrity of alpine meadow ecosystems.The rapid identification of frost mounds and subsequent analysis of their spatial distribution patterns are crucial for monitoring their formation and developmental dynamics and for elucidating the underlying mechanisms.This study focused on the Meiren Grassland in Gannan Prefecture.We employed unmanned aerial vehicle(UAV)photogrammetry to acquire high-resolution visible-light imagery(RGB)and Normalized Digital Surface Model(nDSM)data and subsequently implemented object-based image analysis(OBIA)integrating several machine learning classifiers—namely Bayesian classifier(Bayes),Decision Tree(DT),K-Nearest Neighbors(KNN),Random Forest(RF),and Support Vector Machine(SVM)—to achieve refined frost mound identification.The derived positional information facilitated a comprehensive analysis of frost mound spatial distribution patterns.Our results demonstrated that utilizing fused RGB and nDSM data as input sources,combined with multi-scale segmentation optimized through iterative comparisons(yielding an optimal segmentation scale of 19)and feature space optimization algorithms(selecting an optimal combination of 11 features),the Support Vector Machine(SVM)classifier delivered superior performance.During both the growing season and non-growing season,SVM achieved significantly higher Overall Accuracy(OA=89.49％,90.19％)and Kappa coefficients(74.15％,69.73％)compared to the other classifiers(KNN:OA=87.03％,88.99％;Bayes:OA=86.93％,89.94％;RF:OA=85.35％,87.61％;DT:OA=82.99％,87.60％).Point pattern analysis based on Ripley's L(r)function revealed that frost mounds extracted using all five classification methods exhibited significant clustered distributions in both seasons.Crucially,the optimal SVM method showed distinct scale-dependent patterns:During the growing season,frost mounds displayed a uniform distribution at scales<0.45 m,a random distribution between 0.45 m and 0.46 m,and a clustered distribution at scales>0.46 m.Conversely,during the non-growing season,a uniform distribution was observed at scales<2.13 m,a random distribution between 2.13 m and 2.57 m,and a clustered distribution at scales>2.57 m.The methodology proposed in this study—fusing UAV-derived RGB and nDSM data with object-based machine learning classification—proved highly effective for the refined identification of frost mounds(OA>89％)and multi-scale spatial pattern analysis.This approach provided high-precision technical support for dynamic monitoring of frost mounds within alpine meadows,quantitative assessment of degradation mechanisms,and optimization of mitigation strategies in the context of climate change.