Mapping multi-layered mangroves from multispectral, hyperspectral, and LiDAR data

Abstract

Understanding species distribution and canopy structure of mangrove forests is imperative for flora and fauna conservation in mangrove habitats. However, most mangrove studies focused on the top canopy layer without exploring the vertical structure of mangroves. This paper presents multi-layered mangrove mapping which considered both overstory and understory detection and species classification using multispectral WorldView-3 (WV-3) data, airborne hyperspectral images (HSI), and LiDAR point cloud. First, LiDAR returns were stratified into the overstory and understory by analyzing the profile of return height, which helped understand the vertical structure of the mangrove stands. Second, three classification algorithms Random Forest (RF), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) were compared by applying WV-3, HSI, LiDAR data, and their combinations to map seven vegetative species. Feature selection was conducted to identify important features and the optimal feature size prior to classification tasks. The measured and estimated understory canopy heights reached a high correlation coefficient of 0.71, which demonstrated the effectiveness of using LiDAR data and the proposed procedure to stratify multi-layered canopies. The combined HSI and LiDAR data produced satisfactory results by the three classifiers with overall accuracy (OA) varying from 0.86 to 0.88. And the species was also accurately mapped by integrating WV-3 and LiDAR data using both RF and SVM algorithms with OA attaining between 0.84 and 0.86. The results of this study highlight that (1) LiDAR data provided superior information to map the vertical structure of multi-layered mangroves, which provided valuable information to classify single-layered and dual-layered Kandelia obovata with understory beneath; (2) the combination of spectral and LiDAR features improved mangrove species classification; (3) and species mapping results derived from combined datasets appeared to be more influential by LiDAR features when using RF and SVM, but spectral features played a more important role in CNN.