4:00 PM - 4:15 PM
[ACG37-20] Quantifying changes in biomass density with Sentinel-2A images
Keywords:biomass mapping, forest degradation, sentinel-2A
Activities such as shifting cultivation, cattle grazing, charcoal making and extraction of wood are affecting tropical dry forest, causing disturbances, and affecting ecosystem services. In this study, we use data from forest sampling plots and Sentinel-2A images and derivatives (vegetation index and texture) to find out how those images can capture variations in biomass levels.
We utilized data from 22 sampling plots in tropical dry forest in the Ayuquila watershed, Jalisco, Mexico. The biomass density of those plots varied from very low to relatively high (2-65 Mg/ha), as a result of both differing biophysical conditions and human induced degradation. These plots had a dimension of 500 m2 with 12.6 m of radius, surveyed during 8 days from 9 - 16, May 2022, with measurements for every tree with a diameter of above 5 cm. These measurements include diameter at breast height (DBH), height (m), and canopy cover (%). Aboveground biomass (AGB) data were estimated using an allometric equation (equation 1) that was originally designed for the tropical dry forest in the same state (Martinez-Yrizar et al., 1992).
Log10(AGB) = -0.7590 + 0.9011*Log10(BA)+0.5715*log10(WSG) +0.5654*log10 (height) (1)
in which, AGB is in unit of kg, basal area (BA) in cm2, and height in m. The wood specific gravity (WSG) by species was obtained from database of Zanne et al., (2009). The allometric equation was estimated with 191 samples which has a multiple correlation coefficient of 0.96 and a standard error estimate of 0.16 (Martinez-Yrizar et al., 1992).
We obtained Sentinel-2A images from Google Earth Engine archive which were orthorectified and radiometrically calibrated with the data acquisition date of 15 of march, 2022. The Sentinel-2A bottom of atmosphere reflectance images were atmospherically corrected by Copernicus scihub using the sen2cor algorithm. Texture data of Grey Level Cooccurrence Matrix (GLCM) were calculated based on the band of vegetation index (NDVI) including mean, variance, second moment, entropy, correlation, dissimilarity, contrast, and homogeneity (Haralick et al., 1973).
We considered Sentinel-2A band4 (RED), band8 (NIR), NDVI, and the 8 texture bands as the independent variables to build the biomass model. We extracted the values of the raster bands for each sampling plot. We then used a multiple linear regression to test how the field-measured structural variables including canopy cover, biomass or stem density can be predicted with the spectral and textural variables derived from Sentinel-2A image. We found the linear model could explain only about 45% of the variance in the biomass (with adjusted R-square of 0.454), with a relative large RMSE of 12 Mg/ha. As for the variables, the RED and NIR bands turned out to be significant at 0.05 level, and the NDVI, the texture bands of mean and variance were also significant but at a less degree (0.1 level). This experiment shows that biomass variance could be modeled with Sentinel-2A images but only to a certain degree. In the future experiment, we would like to try using SAR images with both C-and L- band.
References
Haralick, R. M., Shanmugam, K., & Dinstein, I. (1973). Textural features for image classification. Proceedings of the Institute of Electrical and Electronics Engineers Transactions on Systems, Man and Cybernetics, 3, 610–621.
Martinez-Yrizar, A., Sarukhan, J., Perez-Jimenez, A., Rincon, E., Maass, J. M., Solis-Magallanes, A., et al. (1992). Above-ground Phytomass of a Tropical DeciduousForest on the Coast of Jalisco, México. J. Trop. Ecol. 8, 87–96. doi:10.1017/S0266467400006131
Acknowledgement
The first author would like to thank the Direccion General de Asuntos del Personal Académico (DEGAPA) at the Universidad Nacional Autónoma de México (UNAM) for the financial support.
We utilized data from 22 sampling plots in tropical dry forest in the Ayuquila watershed, Jalisco, Mexico. The biomass density of those plots varied from very low to relatively high (2-65 Mg/ha), as a result of both differing biophysical conditions and human induced degradation. These plots had a dimension of 500 m2 with 12.6 m of radius, surveyed during 8 days from 9 - 16, May 2022, with measurements for every tree with a diameter of above 5 cm. These measurements include diameter at breast height (DBH), height (m), and canopy cover (%). Aboveground biomass (AGB) data were estimated using an allometric equation (equation 1) that was originally designed for the tropical dry forest in the same state (Martinez-Yrizar et al., 1992).
Log10(AGB) = -0.7590 + 0.9011*Log10(BA)+0.5715*log10(WSG) +0.5654*log10 (height) (1)
in which, AGB is in unit of kg, basal area (BA) in cm2, and height in m. The wood specific gravity (WSG) by species was obtained from database of Zanne et al., (2009). The allometric equation was estimated with 191 samples which has a multiple correlation coefficient of 0.96 and a standard error estimate of 0.16 (Martinez-Yrizar et al., 1992).
We obtained Sentinel-2A images from Google Earth Engine archive which were orthorectified and radiometrically calibrated with the data acquisition date of 15 of march, 2022. The Sentinel-2A bottom of atmosphere reflectance images were atmospherically corrected by Copernicus scihub using the sen2cor algorithm. Texture data of Grey Level Cooccurrence Matrix (GLCM) were calculated based on the band of vegetation index (NDVI) including mean, variance, second moment, entropy, correlation, dissimilarity, contrast, and homogeneity (Haralick et al., 1973).
We considered Sentinel-2A band4 (RED), band8 (NIR), NDVI, and the 8 texture bands as the independent variables to build the biomass model. We extracted the values of the raster bands for each sampling plot. We then used a multiple linear regression to test how the field-measured structural variables including canopy cover, biomass or stem density can be predicted with the spectral and textural variables derived from Sentinel-2A image. We found the linear model could explain only about 45% of the variance in the biomass (with adjusted R-square of 0.454), with a relative large RMSE of 12 Mg/ha. As for the variables, the RED and NIR bands turned out to be significant at 0.05 level, and the NDVI, the texture bands of mean and variance were also significant but at a less degree (0.1 level). This experiment shows that biomass variance could be modeled with Sentinel-2A images but only to a certain degree. In the future experiment, we would like to try using SAR images with both C-and L- band.
References
Haralick, R. M., Shanmugam, K., & Dinstein, I. (1973). Textural features for image classification. Proceedings of the Institute of Electrical and Electronics Engineers Transactions on Systems, Man and Cybernetics, 3, 610–621.
Martinez-Yrizar, A., Sarukhan, J., Perez-Jimenez, A., Rincon, E., Maass, J. M., Solis-Magallanes, A., et al. (1992). Above-ground Phytomass of a Tropical DeciduousForest on the Coast of Jalisco, México. J. Trop. Ecol. 8, 87–96. doi:10.1017/S0266467400006131
Acknowledgement
The first author would like to thank the Direccion General de Asuntos del Personal Académico (DEGAPA) at the Universidad Nacional Autónoma de México (UNAM) for the financial support.