Addressing the small data problem in forestry: self-supervised learning for aboveground biomass estimation

Abstract Forest aboveground biomass (AGB) and its sub-components (wood, bark, branch foliage) are key elements of forest inventories, informing timber harvesting, carbon accounting, and forest management. Airborne laser scanning (ALS) is commonly used to estimate total and component AGB through the use of metrics (i.e. statistical summaries of point cloud data) across large areas. Recent research has developed deep neural networks (DNNs) that bypass the need for ALS metrics, generating AGB estimates directly from the point cloud. While DNNs typically provide improved AGB estimation compared to conventional models such as random forest (RF), they are limited by data availability. Self-supervised learning (SSL) is a DNN modelling framework that can alleviate large dataset requirements. This study investigates SSL for AGB estimation using ALS data from three perspectives: (i) performance of SSL compared to conventional modelling; (ii) effect of dataset size on SSL performance; (iii) generalization of SSL to out-of-domain data. We implemented an SSL workflow using ALS data (40 points/m2) in a mixedwood boreal forest in Ontario, Canada. The workflow involved pretraining an octree convolutional neural network (OCNN) on 500 000 unlabelled point clouds and fine-tuning OCNN on a smaller ground plot dataset (n = 244 plots). We developed a novel pretext task for OCNN pretraining using ALS metrics as multioutput regression targets to provide the model with prior information about forest structure. Fine-tuning OCNN yielded superior performance for total and component AGB estimation on average (R2 = 0.78, rRMSE = 25.60%) compared to training OCNN from scratch (R2 = 0.73, rRMSE = 28.33%), and RF using ALS metrics (R2 = 0.75, rRMSE = 26.69%). When using incrementally smaller datasets for fine-tuning, we found that SSL outperformed RF in situations with fewer than 60 plots, reducing total AGB relative RMSE by 4 percentage points (6 Mg/ha). The pretrained OCNN also generalized effectively when fine-tuned on an out-of-domain dataset with different forest structure and species composition. This study provides an early demonstration of SSL applied to ALS data for forest attribute modelling and can serve as a foundation for further use of SSL in research and practice.