Forest Aboveground Biomass Estimation Using Satellite Imagery and Random Forest Regression Model

Accurate assessment of forest above-ground biomass is essential for sustainable forest management. Estimation of forest biomass is necessary for studies such as estimation of greenhouse gases, carbon stored in forest resources and climate change models. Also, the forest biomass represents the production rate per unit area. The optical image data of Sentinel-2 satellite was used to estimate the above-ground biomass of the forest in the area of 285 hectares of the forests in Ilam province. 124 square-shaped sample plots with a 20×20 m dimension were located on the ground using a cluster method. Some characteristics of a total of 508 trees (both single stems and coppice forms), including the major and minor crown diameters were measured within each sample plot. Depending on whether the trees are single stem and multi-stem clumps, suitable allometric equations were used to calculate the above-ground biomass based on the measured characteristics. Finally, the total above-ground biomass was calculated for all trees in each sample plot. In order to estimate the above-ground biomass, MSI sensor images of Sentinel 2 satellite were used at the level of L2A corrections. Using spectral ratios, vegetation indices were calculated. In the next step, the corresponding spectral values of the sample plots were extracted from the main bands, and vegetation indices. A random forest regression model was used to estimate forest above-ground biomass. 70% of the samples were used for training the model, and the models were validated using the remaining 30% of the data. The results with R2=0.80 and RMSE=28.70 t/ha showed the acceptable performance of model for estimating the above-ground biomass of the forest. By using the Gini statistic, it was shown that RVI, GNDVI, NDVI, and DVI vegetatuin inices played a larger role in the estimation of biomass.

 Accurate assessment of forest above-ground biomass is essential for sustainable forest management. Estimation of forest biomass is necessary for studies such as the estimation of greenhouse gases, carbon stored in forest resources, and climate change models. Also, the forest biomass represents the production rate per unit area. Estimating forest biomass through direct measurements and cutting and weighing trees in the forests provides an accurate estimate of biomass, but it is a destructive, difficult, and time-consuming method. Therefore, the use of remote sensing methods is very important in the estimation of biomass.

The optical image data of the Sentinel-2 satellite was used to estimate the forest above-ground biomass in the area of 285 hectares of the forests in Ilam province. 124 square-shaped sample plots with a 20×20 m dimension were located on the ground using a cluster sampling strategy. Some characteristics of a total of 508 trees (both single stems and coppice forms), including the major and minor crown diameters were measured within each sample plot. Depending on whether the trees are single-stem or multi-stem clumps, suitable allometric equations were used to calculate the above-ground biomass based on the measured characteristics. Finally, the total above-ground biomass was calculated for all trees in each sample plot. In order to estimate the above-ground biomass, MSI sensor images of the Sentinel 2 satellite were used at the level of L2A corrections. Using spectral ratios, vegetation indices were calculated. In the next step, the corresponding spectral values of the sample plots were extracted from the original bands and vegetation indices. The correlation coefficient between the values of the original bands and vegetation indices with the amount of biomass calculated from the allometric equations in the sample plots was investigated. A random forest regression model was used to estimate forest above-ground biomass. 70% of the samples were used for training the model, and the models were validated using the remaining 30% of the data.

 The results of the descriptive statistics of above-ground forest biomass measured in 120 sample plots which were calculated using allometric equations showed that the lowest biomass in the sample plots is 0.61 and the highest is 268.88 tons per hectare. The average above-ground biomass per tree was measured as 657.6 and 231.2 kg in the single and multi-stemmed trees, respectively. The results of the correlation analysis of biomass with the investigated variables showed that among the main bands of the sensor, the red wavelength has the highest correlation (0.402) with biomass due to the high chlorophyll absorption of green plants in this wavelength. Among the vegetation indices investigated in the research, RVI and NDVI indices have the highest correlation with the forest above-ground biomass with a correlation coefficient of 0.529 and 0.525, respectively. The results of random forest regression analysis to estimate the forest above-ground biomass with R2=0.80, RMSE=28.70 t/ha show the acceptable performance of the model for estimating the above-ground biomass of the forest. Since in this research, the amount of forest above-ground biomass of the sample plots is calculated based on allometric equations in a part of Zagros forests; but these equations are not exactly related to the studied area, part of the model error can be due to this reason. By using the Gini statistic, it was shown that RVI, GNDVI, NDVI, and DVI vegetation indices played a larger role in the estimation of biomass. RVI, NDVI, and DVI indices are calculated using red and near-infrared bands, and since they are influenced by the photosynthetic activity of plants, they are very important in estimating the amount of biomass. GNDVI, which is calculated using green and near-infrared bands, is an indicator of the level of greenness or photosynthetic activity of the plant and is highly sensitive to changes in the chlorophyll content of plants.

 The results of forest above-ground biomass estimation using Sentinel 2 satellite images and random forest regression method showed that using the non-parametric method of the random forest regression model, which performs a large number of uncorrelated models; it has an acceptable ability to estimate forest biomass. Also, the findings showed that vegetation indices are more important in the process of forest above-ground biomass estimation model than Sentinel 2 original bands. The findings of the present research provide the possibility for the managers of Zagros forests to estimate the forest above-ground biomass and provide the basis for sustainable forest management strategies.