Comparison of optimized object-based RF-DT algorithm and SegNet algorithm for classifying Karst wetland vegetation communities using ultra-high spatial resolution UAV data

Abstract

Karst wetlands have the characteristics of small scale and poor stability. At present, the wetland is being severely damaged and its area is seriously degraded, and the accurate identification of vegetation communities is very important for the rapid assessment and management of karst wetland. In this paper, Huixian Karst National Wetland Park, located in Guilin city, China, was taken as the study area, the digital orthophoto map (DOM) and digital surface model (DSM) of UAV images were selected as the data sources, and the vegetation communities of karst wetland were classified by using the object-based Random Forest (RF)-Decision Tree (DT) algorithm and SegNet algorithm. When the object-based RF algorithm and SegNet algorithm were used for coarse classification of karst vegetation, the parameters (mtry, ntree) of the object-based RF algorithm were optimized, and the data dimensionality reduction and RFE variable selection algorithm were used for selecting feature, and the single-class SegNet model was integrated based on the soft voting method to improve the applicability of vegetation classification in karst wetlands. In the classification of vegetation communities in karst wetlands, the optimized object-based RF-DT algorithm were used to extract the vegetation communities in the Areas A, B, and C. The statistical analysis of the importance of the feature variables (spectral features, texture features, geometric features, and position features) of various types of land cover in the three areas was carried out to explore the optimal classification variables of various types of vegetation. The results showed that: (1) the optimized object-based RF algorithm performed better than the SegNet algorithm in classifying karst vegetation at 95% confidence level during the coarse classification. The average accuracy of wetland vegetation was improved by 1.06–13.58%; (2) the object-based RF-DT algorithm had high classification ability for the karst wetland vegetation community, with overall accuracy and kappa coefficient above 0.85; and that (3) although geometric features accounted for the largest proportion (52.2%) in the classification of bermudagrass, water hyacinth, lotus, linden and other vegetation, texture features accounted for the highest proportion of 56.3% in the classification of vegetation whose importance was more than 90.