Canadian Forest Service Publications

Object-based random forest modelling of aboveground forest biomass outperforms a pixel-based approach in a heterogeneous and mountain tropical environment. 2019. Silveira, E.M.O., S. H..G. Silva, F.W. Acerbi-Júnior, M.C. Carvalho, L.M.T. Carvalho, J.R.S. Scolforo, Wulder, M.A. Int J Appl Earth Obs Geoinformation, 78, pp 175-188.

Year: 2019

Issued by: Pacific Forestry Centre

Catalog ID: 40121

Language: English

Availability: PDF (request by e-mail)

Available from the Journal's Web site.
DOI: 10.1016/j.jag.2019.02.004

† This site may require a fee

Mark record

Abstract

The Brazilian Atlantic Forest is a highly heterogeneous biome of global ecological significance with high levels of terrestrial carbon stocks and aboveground biomass (AGB). Accurate maps of AGB are required for monitoring, reporting, and modelling of forest resources and carbon stocks. Previous research has linked plot-level AGB with environmental and remotely sensed data using pixel-based approaches. However, few studies focused on investigating possible improvements via object-based image analysis (OBIA) including terrain related data to predict AGB in topographically variable and mountainous regions, such as Atlantic forest in Minas Gerais, Brazil. OBIA is expected to reduce known uncertainties related to the positional discrepancy between the image and field data and forest heterogeneity, while terrain derivatives are strong predictors in forest ecosystems driving forest biomass variability. In this research, we compare an object-based approach to a pixel-based method for modeling, mapping and quantifying AGB in the Rio Doce basin, within the Brazilian Atlantic Forest biome. We trained a random forest (RF) machine learning algorithm using environmental, terrain, and Landsat Thematic Mapper (TM) remotely sensed imagery. We aimed to: (i) increase the precision of the AGB estimates; (ii) identify optimal variables that fit the best model, with the lowest root mean square error (RMSE, Mg/ha); (iii) produce an accurate map of the AGB for the study area, and subsequently (iv) describing the AGB spatial distribution as a function of the selected variables. The RF object-based model notably improved the AGB prediction by reducing the mean absolute error (MAE) from 28.64 to 20.95%, and RMSE from 33.43 to 20.08 Mg/ha, and increasing the R² (from 0.57 to 0.86) by using a combination of selected remote sensing, environmental, and terrain variables. Object-based modelling is a promising alternative to common pixel-based approaches to reduce AGB variability in topographically diverse and heterogeneous environments. Investigation of mapped outcomes revealed a decreasing AGB from west towards the east region of the Rio Doce Basin. Over the entire study area, we map a total of 195,799,533 Mg of AGB, ranging from 25.52 to 238 Mg/ha, following seasonal precipitation patterns and anthropogenic disturbance effects. This study provided reliable AGB estimates for the Rio Doce basin, one of the most important watercourses of the globally important Brazilian Atlantic Forest. In conclusion, we highlight that OBIA is a better solution to map forest AGB than the pixel-based traditional method, increasing the precision of AGB estimates in a heterogeneous and mountain tropical environment.

Plain Language Summary

The Brazilian Atlantic Forest is a highly heterogeneous biome of global ecological significance with high levels of terrestrial carbon stocks and aboveground biomass (AGB). Accurate maps of AGB are required for monitoring, reporting, and modelling of forest resources and carbon stocks. Previous research has linked plot-level AGB with environmental and remotely sensed data using pixel-based approaches. However, few studies focused on investigating possible improvements via object-based image analysis (OBIA) including terrain related data to predict AGB in topographically variable and mountainous regions, such as Atlantic forest in Minas Gerais, Brazil. OBIA is expected to reduce known uncertainties related to the positional discrepancy between the image and field data and forest heterogeneity, while terrain derivatives are strong predictors in forest ecosystems driving forest biomass variability. In this research, we compare an object-based approach to a pixel-based method for modeling, mapping and quantifying AGB.