Multi- and hyperspectral remote-sensing retrieval of floodplain-forest aboveground biomass using machine learning

Wednesday, 17 December 2014
Anthony M Filippi, Texas A & M Univ, College Station, TX, United States, Inci Guneralp, Texas A & M University, Geography, College Station, TX, United States and Jarom Randall, Texas A & M University, College Station, TX, United States
Forests within dynamic floodplain landscapes, such as meandering-river landscapes, are composed of uneven-aged trees and entail high spatial variability, which results from intersecting hydrological, fluvial, and ecological processes. Floodplain forests are an important carbon sink relative to other terrestrial ecosystems and thus serve a critical role in the global carbon cycle. Accurate, quantitative aboveground biomass (AGB) retrieval within floodplain forests is urgently needed for improved carbon-pool estimates in such areas and enhanced process understanding of river-floodplain biomorphodynamics. We perform remote AGB retrieval for a meander-bend bottomland hardwood forest, based on utilization of stochastic gradient boosting (SGB), multivariate adaptive regression splines (MARS), and Cubist algorithms and multi- and hyperspectral image-based data sets. For multispectral experiments, we use 30-m and 10-m image bands (Landsat 7 ETM+ and SPOT 5, respectively) and ancillary input vectors; for hyperspectral-based experiments, we use 30-m Hyperion bands and other input variables. Results indicate that for both the multispectral and hyperspectral experimental trials, SGB- and MARS-derived AGB are significantly more accurate than Cubist estimates. (Cubist is used for U.S. national-scale forest biomass mapping.) For the multispectral results, across all data-experiments and algorithms, at 10-m spatial resolution, SGB gives the most accurate estimates (RMSE = 22.49 tonnes/ha; coefficient of determination (R2) = 0.96) when geomorphometric data are also included. For 30-m multispectral data trials, MARS performs the best (RMSE = 29.2 tonnes/ha; R2 = 0.94) when image-derived data are also incorporated. For the hyperspectral experiments, the most accurate MARS- and SGB-based retrievals have R2 of 0.97 and 0.95, respectively; the most accurate Cubist AGB retrieval has R2 of 0.85. MARS and SGB AGB are not significantly different though for the hyperspectral experiments. The SGB and MARS approaches may enable improved operational AGB monitoring across scales and environments.