Aboveground biomass inversion based on Sentinel-2 remote sensing images in Chongli district

Based on Sentinel-2 remote sensing image data, combined with the actual forest sample plot data, we proposed a new idea of inversion of forest above-ground biomass in Chongli district as an example.

Based on Sentinel-2 remote sensing image data of Chongli district, Hebei province, in July 2021, and 71 forest sample plots of Chongli district in June and August 2021, the measured data were used. Using the diameter at breast height and tree height from the measured data, the measured biomass of each plot was calculated according to the formula for calculating forest biomass in Hebei province. The remote sensing data were pre-processed with SNAP and ENVI software for resampling and cropping to extract the original image bands and calculate remote sensing factors such as vegetation index, texture factor and tassel cap index. Pearson correlation analysis was performed to filter the remote sensing factors and optimize the selection of texture factors by matching the best texture window size. Three algorithms of multiple linear regression, BP neural network and random forest were used to model the AGB of Chongli district, respectively. The model accuracy was evaluated using R² and RMSE, and the optimal model was selected for biomass inversion and biomass spatial distribution mapping.

1) In the selection of remote sensing factors, in addition to the conventional green band, red band and two vegetation red edge bands and vegetation indices DVI, SAVI and EVI, the mean value of texture factor and the brightness and greenness of tassel hat index also played an important role in the establishment of biomass inversion model, and the selection of texture factor window size also affected the accuracy of the final model; 2) The accuracy of all three models met the requirements of biomass inversion, with the best random forest model, the second best multiple linear regression model, and the lowest accuracy of BP neural network model, but the accuracy of the BP model improved after the ten-rule cross-validation method, and the R² of the optimal random forest model reached 0.843; 3) After the inversion of the optimal model, the distribution of AGB in the Chongli area mainly ranged from 50-200 mg·hm^-2, concentrated in the western ring of mountains, with obvious spatial heterogeneity.

The inversion of forest biomass using Sentinel-2 remote sensing images has high accuracy. With the addition of vegetation index, tassel cap index and texture factor, the model effect showed an increasing trend, and the window size selection of texture factor had an important influence in the remote sensing inversion of forest biomass.