Reducing classification error of grassland overgrowth by combing low-density lidar acquisitions and optical remote sensing data

Abstract

Mapping structural changes in vegetation dynamics has, for a long time, been carried out using satellite images, orthophotos and, more recently, airborne lidar acquisitions. Lidar has established its position as providing accurate material for structure-based analyses but its limited availability, relatively short history, and lack of spectral information, however, are generally impeding the use of lidar data for change detection purposes. A potential solution in respect of detecting both contemporary vegetation structures and their previous trajectories is to combine lidar acquisitions with optical remote sensing data, which can substantially extend the coverage, span and spectral range needed for vegetation mapping. In this study, we tested the simultaneous use of a single low-density lidar data set, a series of Landsat satellite frames and two high-resolution orthophotos to detect vegetation succession related to grassland overgrowth, i.e. encroachment of woody plants into semi-natural grasslands. We built several alternative Random Forest models with different sets of variables and tested the applicability of respective data sources for change detection purposes, aiming at distinguishing unchanged grassland and woodland areas from overgrown grasslands. Our results show that while lidar alone provides a solid basis for indicating structural differences between grassland and woodland vegetation, and orthophoto-generated variables alone are better in detecting successional changes, their combination works considerably better than its respective parts. More specifically, a model combining all the used data sets reduces the total error from 17.0% to 11.0% and omission error of detecting overgrown grasslands from 56.9% to 31.2%, when compared to model constructed solely based on lidar data. This pinpoints the efficiency of the approach where lidar-generated structural metrics are combined with optical and multitemporal observations, providing a workable framework to identify structurally oriented and dynamically organized landscape phenomena, such as grassland overgrowth.