Forest Aboveground Biomass Estimation Using Multi-Source Remote Sensing Data in Temperate Forests

Forests are a crucial part of global ecosystems. Accurately estimating aboveground biomass (AGB) is important in many applications including monitoring carbon stocks, investigating forest degradation, and designing sustainable forest management strategies. Remote sensing techniques have proved to be a cost-effective way to estimate forest AGB with timely and repeated observations. This dissertation investigated the use of multiple remotely sensed datasets for forest AGB estimation in temperate forests. We compared the performance of Landsat and lidar data—individually and fused—for estimating AGB using multiple regression models (MLR), Random Forest (RF) and Geographically Weight Regression (GWR). Our approach showed MLR performed similarly to GWR and both were better than RF. Integration of lidar and Landsat inputs outperformed either data source alone. However, although lidar provides valuable three-dimensional forest structure information, acquiring comprehensive lidar coverage is often cost prohibitive. Thus we developed a lidar sampling framework to support AGB estimation from Landsat images. We compared two sampling strategies—systematic and classification-based—and found that the systematic sampling selection method was highly dependent on site conditions and had higher model variability. The classification-based lidar sampling strategy was easy to apply and provides a framework that is readily transferable to new study sites. The performance of Sentinel-2 and Landsat 8 data for quantifying AGB in a temperate forest using RF regression was also tested. We modeled AGB using three datasets: Sentinel-2, Landsat 8, and a pseudo dataset that retained the spatial resolution of Sentinel-2 but only the spectral bands that matched those on Landsat 8. We found that while RF model parameters impact model outcomes, it is more important to focus attention on variable selection. Our results showed that the incorporation of red-edge information increased AGB estimation accuracy by approximately 6%. The additional spatial resolution improved accuracy by approximately 3%. The variable importance ranks in the RF regression model showed that in addition to the red- edge bands, the shortwave infrared bands were important either individually (in the Sentinel-2 model) or in band indices. With the growing availability of remote sensing datasets, developing tools to appropriately and efficiently apply remote sensing data is increasingly important