A Point Cloud Transformer-Based Regression Approach for Estimating Above-Ground Biomass from 3D Point Clouds

Abstract

From the perspective of global food security, the estimation of crop biomass and yield is a critical process. Traditionally, plant height is manually measured using field samples for biomass and yield estimation, and plant samples are cut, dried, and weighed. This labor-intensive process can be made more efficient through advanced technologies integrated into modern agricultural machinery. One such technology, light detection and ranging (LiDAR), provides nondestructive 3D point clouds with millimeter-level accuracy. However, processing this data is highly challenging due to the complex canopy architecture of cereal crops. Therefore, deep learning architectures can be used to effectively analyze 3D point cloud data. In this context, a deep learning-based regression approach was proposed to estimate above-ground biomass from 3D point cloud data. This study employed a point cloud transformer (PCT) architecture. Growth Stage-PCT (GS-PCT) models were developed by integrating growth stage information into the proposed Small, Base, and Deep variants of the PCT architecture. According to the experimental results, when both growth stages were evaluated together, the proposed GS-PCT-Small model achieved the best performance, with an R² of 0.95, an RMSE of 88.60 g m⁻², an MAE of 69.34 g m⁻², and a MAPE of 13.69%. During the vegetative stage, the PCT-Small model achieved an RMSE of 58.52 g m⁻², while during the flowering stage, the GS-PCT-Small model achieved an RMSE of 111.55 g m⁻². The findings indicate that the models achieved high predictive performance for above-ground biomass estimation from 3D point clouds.