基于机器学习模型的大比例尺土地自然类型分布预测OACHSSCD

The survey and mapping of land natural types serve as the foundation for natural resource management and the optimization of territorial space,and are of great significance for the rational utilization of land resources and ecological protection.However,at present,there is still a notable lack of large-scale regional mapping and high-precision automated classification methods,which hampers the ability to meet the demands of refined management of policy decision-makers.This study aims to integrate field surveys with machine learning to establish a novel high-precision classification method for land natural types,and verify it in two representative case areas with significant physiographical differences in China,namely the field observation station in the lower reaches of the Tarim River in Xinjiang and the field observation station in Sanya,Hainan.We adopted methods such as field unmanned aerial vehicle surveying,vegetation quadrat investigation,and soil property detection to construct a"vegetation-soil-landform"classification system for land natural types.And by applying the Random Forest model,the automatic classification of land natural types is achieved based on multi-source environmental variables,aiming to provide a scientific basis for the utilization and management of land resources.The results show that:(1)In the typical study areas of Xinjiang and Hainan,55 and 25 land natural types were respectively classified,and the prediction accuracy rates of the Random Forest classification model reached 87.47％—89.48％and 95.92％—97.83％,respectively;(2)Soil water content and elevation were the dominant environmental factors influencing land natural type differentiation in the study areas of Xinjiang and Hainan;(3)In areas lacking field-measured data,using publicly available remote sensing products as substitutes resulted in only a 1.1％—5.3％decrease in model accuracy.This study has constructed a transferable"field survey+machine learning"framework for large-scale land natural type classification.Based on the naming method of"vegetation-soil-landform"and the machine learning classification model,it can achieve high-precision regional classification of land natural types.Even in natural areas where it is difficult to obtain measured data and the accessibility is low,remote sensing products can be used for classification and mapping with guaranteed accuracy.This makes the Random Forest model have great application potential in large-scale land natural type classification and mapping research.In the future,it can be extended to a wider range of pilot studies in combination with major national demands,promoting the practical application of land natural type mapping in natural resource management,and providing precise and efficient data support for territorial space planning and ecological protection and restoration.