Research on Unsupervised Vegetation Remote Sensing Mapping Method Based on Sample Migration

Objectives: Vegetation classification and mapping are of great significance to developing ecological environmental protection. Supervised classification is the most widely used method for vegetation classification and mapping because it can generate accurate classification results. However, most of the current vegetation classification and mapping methods rely on single-phase field data, and field sampling often requires a lot of manpower and material resources, so it is difficult to realize long-time sequence dynamic vegetation classification and mapping only by field sample data. Methods: This study proposes an unsupervised classification method for automatically obtaining reliable samples based on historical vegetation classification map datasets. Using the 1:1 000 000 Chinese Vegetation Atlas dataset as prior knowledge, a model for optimizing and migrating large-label samples for vegetation types is proposed. Multi-source data of the same or similar temporal phase as the existing vegetation classification maps are used to construct a feature set for local sample clustering optimization and global sample hierarchical Gaussian mixture optimization, which replaces the manual selection of training samples, and then obtains a usable sample set. On this basis, the invariant region sample migration is carried out by combining the results of the long-time-series Landsat vegetation change detection. The migrated samples are used for multi-temporal vegetation classification mapping, to quickly obtain the dynamic mapping results of long time series vegetation classification. Results: The experiment selected Arhorqin Banner as the key study area, and completed the regional multi-temporal natural vegetation classification and mapping from 2005 to 2022, with the overall accuracy of the classification better than 88%, and the Kappa greater than 0.80. The natural vegetation in the study area is dominated by grasslands and forests and scrubs are mainly distributed in the northwestern part of the study area, with temperate tufted grassland dominating the grasslands, and the temperate tufted grassland dominated by the extensive needlegrass grassland. From 2005 to 2022, the degree of shrinkage of natural vegetation gradually increases from north to south, in which the area of temperate deciduous broad-leaved forests does not change significantly, the most significant change in natural vegetation is located in the southern part of the study area, temperate graminoid grassland, temperate deciduous scrub (dominated by small-leaved mallard scrub) decreases year by year and the area of non-natural vegetation increases year by year. Further analysis of the situation shows that the southern part of the study area is dominated by sandy vegetation, which has been seriously degraded since 2000. Still, since 2010 the study area has started to implement the artificial forage industry, so the natural vegetation area in the southern part of the study area has decreased. The unnatural vegetation area has increased, and the unnatural vegetation area is mostly small round patches, which are more concentrated in the southern part of the study area. Conclusions: The overall accuracy of the classification meets the needs of long-time series vegetation classification mapping, and the mapping effect is more stable. Therefore, the unsupervised sample migration method based on historical large-label vegetation classification maps can to some extent make full use of the existing vegetation classification products, and provide a more convenient way for vegetation classification mapping update, using the geometric features of the historical vegetation classification products, combining with the multi-source remote sensing data of the same or similar time-phase to construct the feature attribute set of the patch-by-patch, and hierarchically optimizing samples of the large-label data of vegetation classification from local to global, to form the optimized and migratable vegetation classification maps. Optimization of samples is carried out, to form a migratable training set of optimized vegetation classification samples. The optimized vegetation classification samples training set is then combined with land use data to carry out multi-temporal natural vegetation classification and mapping, and the results of long time-series natural vegetation classification changes are obtained. This study provides a fast, convenient, lightweight, and reliable mapping method for vegetation remote sensing classification.