A fully-automated approach to land cover mapping with airborne LiDAR and high resolution multispectral imagery in a forested suburban landscape

Abstract

Information on land cover is essential for guiding land management decisions and supporting landscape-level ecological research. In recent years, airborne light detection and ranging (LiDAR) and high resolution aerial imagery have become more readily available in many areas. These data have great potential to enable the generation of land cover at a fine scale and across large areas by leveraging 3-dimensional structure and multispectral information. LiDAR and other high resolution datasets must be processed in relatively small subsets due to their large volumes; however, conventional classification techniques cannot be fully automated and thus are unlikely to be feasible options when processing large high-resolution datasets. In this paper, we propose a fully automated rule-based algorithm to develop a 1m resolution land cover classification from LiDAR data and multispectral imagery.The algorithm we propose uses a series of pixel- and object-based rules to identify eight vegetated and non-vegetated land cover features (deciduous and coniferous tall vegetation, medium vegetation, low vegetation, water, riparian wetlands, buildings, low impervious cover). The rules leverage both structural and spectral properties including height, LiDAR return characteristics, brightness in visible and near-infrared wavelengths, and normalized difference vegetation index (NDVI). Pixel-based properties were used initially to classify each land cover class while minimizing omission error; a series of object-based tests were then used to remove errors of commission. These tests used conservative thresholds, based on diverse test areas, to help avoid over-fitting the algorithm to the test areas.The accuracy assessment of the classification results included a stratified random sample of 3198 validation points distributed across 30 1×1km tiles in eastern Connecticut, USA. The sample tiles were selected in a stratified random manner from locations representing the full range of rural to urban landscapes in eastern Connecticut. The overall land cover accuracy was 93% with accuracies exceeding 90% for deciduous trees, low vegetation, water, buildings, and low impervious cover. Slight confusion occurred between coniferous and deciduous trees; major confusion occurred between water and riparian wetlands; and moderate confusion occurred between medium vegetation and other vegetation classes. The algorithm was robust for the forested suburban landscape of eastern Connecticut, which is typical for much of the northeastern U.S., and the algorithm shows promise for applications in similar landscapes with similar datasets. Further research is needed to test the applicability of the algorithm to more diverse landscapes as well as with different LiDAR and multispectral datasets.