BioPM: Mixer for Point Cloud Based Biomass Prediction

AGB(Above-Ground Biomass) is crucial trait relevant to agricultural production and study. Benefiting from the availability of field point cloud scanned by LiDAR, it's possible to use a non-destructive and high-throughput method for predicting AGB instead of laborious and destructive methods. Inspired by deep learning methods in 3D object detection by grouping point cloud and Mixer structure achieves great performance on 2D computer vision tasks, we propose an end-to-end prediction network BioPM, which combines both advantages based on the upward growth characteristics of wheat. Our BioPM consists of two modules: 1) a feature encoding module to group point cloud as pillars and extract point-wise features of pillars; and 2) a mixer module to extract pillar-wise features and output predictions by using only MLP. Experiments on the public dataset show that our BioPM prediction outperforms non-deep learning SOTA methods and other deep learning methods.