Utilising Space-borne 3D Vegetation Metrics to Improve Land Cover Mapping: A Case Study from A Complex Tropical Landscape

Abstract:

Land cover is an essential variable in understanding the global climate and terrestrial carbon cycle. The most common approach to derive land cover information is through the classification of optical satellite images. In complex landscapes, however, land cover types are often inter-confused with those that are spectrally similar but structurally different. For example, discrimination between different forest types and tree plantations may be challenging with spectral information alone. We examined the potential of vertical vegetation structure acquired by the GEDI (Global Ecosystem Dynamics Investigation), a space-borne waveform lidar sensor installed on the ISS, to improve land cover maps across a complex seasonally dry tropical landscape in Cambodia. We extracted a range of GEDI metrics from Level-1b and Level-2a data products at over 421 locations across 6 known land cover types across the study area. A set of decision trees were developed to predict land cover based on these structural metrics. Once developed, we applied the model to 79,000 GEDI footprints over the area and compared them to an existing land cover dataset derived from the fusion of Sentinel-1 and Sentinel-2 data.

Our first results showed that GEDI ground elevation, ground wave portion, and top canopy height metrics were the best predictors of land cover. The cover type with the highest prediction accuracy was annual crop (User’s accuracy (UA) = 92%; Producer’s accuracy (PA) = 100%). Evergreen forest had the lowest accuracy (UA = 94%; PA = 75%). Nevertheless, the GEDI metrics were able to accurately discriminate the two major types of natural forest, evergreen and deciduous savanna-like forest, and rubber plantation from forests with reliable accuracy.

The comparison of GEDI predicted land cover with the existing land cover information showed the closest agreement with annual crops. The greatest land cover differences occurred within the seasonally-inundated forest and rubber plantation classes.

Our results demonstrate the potential of structural information from GEDI to improve the performance of land cover classifications in complex, heterogeneous landscapes. GEDI data enhance our ability to identify different vegetation types that require distinct management approaches and allow for more effective landscape management practices.