5:15 PM - 6:30 PM
[ACG37-P08] Estimation of forest above ground biomass from PALSAR Data
Keywords:forest aboveground biomass, ALOS, random forest, support vector machine, Landsat-8 OLI
Under climate change and increasingly prominent environmental issues, the estimation on forest carbon stock has become hot research topic. Huge volume and high carbon content of tree is the main carbon storage reservoir of the terrestrial biosphere and plays an important role in the global carbon cycle. Therefore, accurate estimation of forest biomass enables us to evaluate and track the status of forest carbon stocks, and to predict the role and feedback of forests in the regional and global carbon cycle, and to provide an important basis for specifying policies to response climate change.
Remote sensing approach offers considerable contribution to forest monitoring via long-term and repetitive observations over vast area (Mermoz et al 2015). Synthetic aperture radar (SAR) is one of the most promising remote detectors to map the global forest above ground biomass (AGB). Especially, the L-band SAR is widely used to retrieve forest biomass, primarily based on a positive correlation between SAR backscatter and in situ AGB. A few studies have successfully used SAR data to map AGB, such as in Madagascar by Minh et al (2018), in Matang Mangroves, Malaysia b Hamdan et al (2014), in Africa by Bouvet et al (2018). These proved that the L-band SAR had a significant relationship with forest biomass, however, many studies have also proved that when AGB reaches a certain level, the correlation between it and the signal will decrease, which is also called saturation phenomenon. Inevitably, SAR shows poor accuracy in high biomass. In addition, satellite attitudes will also bring considerable noise to the inversion results. Although these noises can be avoided by geometric correction and extraction of areas with better local incident angle conditions, a relatively high-precision inversion result can still be hard to achieve.
The purpose of this study is to try to improve the inversion accuracy through filtering calculations and machine learning methods.
This study takes Hokkaido as the research area, based on ALOS-PALSAR and optical remote sensing data, combined with field survey data of forest resources, extracts a variety of remote sensing data modeling variables, and uses random forest and support vector machine methods to target the forest biomass of different tree species. Noted that there are two sets of variables for modeling, which are optical remote sensing data and SAR data. The forest coverage rate in Hokkaido is over 70%, with abundant vegetation types and species. Naturally distributed species mainly include Pinus densiflora, Abies sachalinensis, Cryptomeria japonica, Picea jezoensis, Fagus crenata, oak, etc., covering broad-leaved forests, coniferous forests, Coniferous and broad-leaved mixed forest. In this study, the sample plot data we used are the forest resource survey data provided by the Japanese Forestry Agency. The field data collection time was from 2004 to 2008, a total of 2907 sample sites were collected, and the area of each sample site was 1 ha. The main investigation factors of the field data of the Japanese Forestry Agency include tree height, tree age, volume per ha, tree types, slope, aspect, etc. In this study, the forest aboveground biomass is calculated by the biomass-accumulation relationship equation (Xu et al, 2007). In order to match the acquisition time of the field sample location as much as possible in time, we selected PALSAR data from 2007 to 2008, with two polarization modes, HH and HV, with a resolution of 25 meters. The accuracy of forest biomass estimation by different modeling methods, different tomographic parameters and different forest types was compared. Finally mapping the forest aboveground biomass in Hokkaido.
Remote sensing approach offers considerable contribution to forest monitoring via long-term and repetitive observations over vast area (Mermoz et al 2015). Synthetic aperture radar (SAR) is one of the most promising remote detectors to map the global forest above ground biomass (AGB). Especially, the L-band SAR is widely used to retrieve forest biomass, primarily based on a positive correlation between SAR backscatter and in situ AGB. A few studies have successfully used SAR data to map AGB, such as in Madagascar by Minh et al (2018), in Matang Mangroves, Malaysia b Hamdan et al (2014), in Africa by Bouvet et al (2018). These proved that the L-band SAR had a significant relationship with forest biomass, however, many studies have also proved that when AGB reaches a certain level, the correlation between it and the signal will decrease, which is also called saturation phenomenon. Inevitably, SAR shows poor accuracy in high biomass. In addition, satellite attitudes will also bring considerable noise to the inversion results. Although these noises can be avoided by geometric correction and extraction of areas with better local incident angle conditions, a relatively high-precision inversion result can still be hard to achieve.
The purpose of this study is to try to improve the inversion accuracy through filtering calculations and machine learning methods.
This study takes Hokkaido as the research area, based on ALOS-PALSAR and optical remote sensing data, combined with field survey data of forest resources, extracts a variety of remote sensing data modeling variables, and uses random forest and support vector machine methods to target the forest biomass of different tree species. Noted that there are two sets of variables for modeling, which are optical remote sensing data and SAR data. The forest coverage rate in Hokkaido is over 70%, with abundant vegetation types and species. Naturally distributed species mainly include Pinus densiflora, Abies sachalinensis, Cryptomeria japonica, Picea jezoensis, Fagus crenata, oak, etc., covering broad-leaved forests, coniferous forests, Coniferous and broad-leaved mixed forest. In this study, the sample plot data we used are the forest resource survey data provided by the Japanese Forestry Agency. The field data collection time was from 2004 to 2008, a total of 2907 sample sites were collected, and the area of each sample site was 1 ha. The main investigation factors of the field data of the Japanese Forestry Agency include tree height, tree age, volume per ha, tree types, slope, aspect, etc. In this study, the forest aboveground biomass is calculated by the biomass-accumulation relationship equation (Xu et al, 2007). In order to match the acquisition time of the field sample location as much as possible in time, we selected PALSAR data from 2007 to 2008, with two polarization modes, HH and HV, with a resolution of 25 meters. The accuracy of forest biomass estimation by different modeling methods, different tomographic parameters and different forest types was compared. Finally mapping the forest aboveground biomass in Hokkaido.