Automated Biomass Estimation Leveraging Instance Segmentation and Regression Models with UAV Aerial Imagery and Forest Inventory Data

The net-zero emission movement has increased the priority of the green energy transition, which is mainly supported by renewable sources. The above-ground biomass (AGB) specifies the amount of mass convertible to energy. Due to its importance, automatic and accurate AGB estimation is required since traditional methods are labor-intensive and impractical for large areas. To address the challenge, this study applies UAV aerial imagery into a processing pipeline, including instance segmentation and regression analysis, to estimate AGB. The process begins with obtaining pixel information for crown areas of tree species in Dulamayo and Tupa forests via instance segmentation. Subsequently, the identified trees’ trunk diameter and height are calculated through a regression analysis process utilizing secondary information, namely global allometric and wood density database. Our experimental results show that crown diameter has a strong $\mathbf{R}^{\mathbf{2}}$ value of $\mathbf{0.74}$ for trunk diameter prediction, while multivariate regression using crown diameter and predicted trunk diameter achieved an $\mathbf{R}^{2}$ value of $\mathbf{0.80}$ for tree height prediction. Further, wood density is inferred by averaging values from the global database for species closely related to coconut (Cocos nucifera), sugar palm (Arenga pinnata), and clove (Syzygium aromaticum). An established AGB allometric model, incorporating trunk diameter, wood density, and tree height, is used to assist AGB calculation due to the absence of AGB ground truth information. The approach aims at reducing uncertainty when predicting AGB. The devised processing pipeline demonstrates the potential application of artificial intelligence in biomass estimation taking place in remote and resource-limited areas.