Estimates of Aboveground Biomass from Texture Analysis of Landsat Imagery

• Published in: Remote sensing (Basel, Switzerland) Vol. 6; no. 7; pp. 6407 - 6422
• Main Authors: Kelsey, Katharine C, Neff, Jason C
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.07.2014
• Subjects: Biomass; carbon; Coefficient of variation; Correlation; Entropy; error; forest; Forest biomass; Landsat; Mathematical models; National forests; Natural resource management; regional biomass map; Remote sensing; San Juan National Forest; Satellite imagery; Spectra; Surface layer; Texture
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs6076407

Maps of forest biomass are important tools for managing natural resources and reporting terrestrial carbon stocks. Using the San Juan National Forest in Southwest Colorado as a case study, we evaluate regional biomass maps created using physical variables, spectral vegetation indices, and image textural analysis on Landsat TM imagery. We investigate eight gray level co-occurrence matrix based texture measures (mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation) on four window sizes (3 3, 5 5, 7 7, 9 9) at four offsets ([1,0], [1,1], [0,1], [1,-1]) on four Landsat TM bands (2, 3, 4, and 5). The map with the highest prediction quality was created using three texture metrics calculated from Landsat Band 2 on a 3 3 window and an offset of [0,1]: entropy, mean and correlation; and one physical variable: slope. The correlation of predicted versus observed biomass values for our texture-based biomass map is r = 0.86, the Root Mean Square Error is 45.6 Mg times ha-1, and the Coefficient of Variation of the Root Mean Square Error is 0.31. We find that models including image texture variables are more strongly correlated with biomass than models using only physical and spectral variables. Additionally, we suggest that the use of texture appears to better capture the magnitude and direction of biomass change following disturbance compared to spectral approaches. The biomass mapping methods we present here are widely applicable throughout the US, as they are based on publically available datasets and utilize relatively simple analytical routines.