Two-step carbon storage estimation in urban human settlements using airborne LiDAR and Sentinel-2 data based on machine learning

Abstract

The quantification of carbon storage (CS) within urban areas has become increasingly crucial for achieving global carbon neutrality. This study proposed a new approach to estimating CS in urban human settlements, building on an existing forestry biomass expansion factor (BEF)-based CS estimation approach, and tested it over Suwon, a city in South Korea. First, a tree canopy cover map was created using high resolution land cover data. The urban tree area ratio was then calculated to estimate the BEF-based CS (Step 1 CS), considering parameters from the national forest inventories. Since urban trees have different growing environments and characteristics than forests, the CS for human settlements (Step 2 CS) was estimated using Step 1 CS as well as structural and spectral information of trees from light detection and ranging (LiDAR) and Sentinel-2 data via machine learning (ML) techniques. The dependent variable of the ML models was the difference between the Step 1 CS and the reference CS, which was calculated using field-measured data and existing allometric equations for urban areas. Using both LiDAR and Sentinel-2 information, the random forest model outperformed the other ML models tested, with an R2 of 0.884, a root mean squared error of 0.432 tC/900 m2, and a lower sampling strategy sensitivity. Feature analysis revealed that incorporating structural and spectral information resulted in a more reliable CS estimation by considering different tree environmental characteristics for each grid cell in the model. This data-driven model utilizes remote sensing and ML and can generate spatially explicit CS maps of large urban areas. It can accurately quantify CS in human settlements and is expected to be easily applicable to other urban green spaces.