Forest aboveground biomass estimation using deep learning data fusion of ALS, multispectral, and topographic data

Estimating forest aboveground biomass (AGB) and its components (wood, branch, bark, foliage) is critical for forest inventories and provides important information for timber harvesting and carbon accounting. Current approaches for modelling forest AGB at the stand scale often employ airborne laser scanning (ALS) data which provide robust AGB estimates. However, in structurally complex forest ecosystems, ALS-based models may not estimate forest biomass and its components with sufficient accuracy. One method to improve ALS-based model performance is through data fusion. Deep neural networks (DNNs) are effective for data fusion because they can combine different data modalities without the need to modify the original data resolution. This study evaluated the effectiveness of a data fusion DNN that combines ALS, multispectral, and topographic data for forest biomass estimation (total and component). We implemented a DNN architecture consisting of three convolutional neural network (CNN) modules: Octree-CNN for ALS data; 1-D CNN for Landsat-8 multispectral data; and 2-D CNN for topographic data. Variants of the DNN architecture combining different input data modalities were trained and tested using sample plots from New Brunswick, Canada (n = 2,336). The model, including all three data modalities, performed best overall for total AGB estimation (R2 = 0.77; RMSE = 28.38 Mg/ha) and explained an additional 2−5% variation in wood, bark, and foliage biomass compared to the ALS-only model. This study demonstrates the effectiveness of a novel data fusion DNN architecture that extracts information directly from input data modalities for improving forest biomass estimates. However, relatively small performance gains should be weighed against computational resources and domain knowledge required to implement and interpret DNNs.