A scalable framework for UAV-based point cloud reconstruction and organ segmentation of field-grown cotton in complex field environments

Abstract

Accurate characterization of plant 3D architecture and semantic parsing of key cotton organs represent an essential prerequisite for crop phenotyping, precision field management, and cultivar breeding and selection. However, the overall 3D structure of field-grown cotton plants is highly complex in field environments. The morphological traits and spatial distribution of various organs are modulated by multiple factors including genetics, environmental conditions, and cultivation management practices, resulting in pronounced phenotypic variation under field conditions. Thus, this study proposed an integrated technical framework for UAV-based 3D reconstruction and organ semantic segmentation tailored for field-grown cotton. High-fidelity 3D point clouds of the crop canopy are generated by coupling an field-adapted low-altitude unmanned aerial vehicle (UAV) acquisition strategy with neural radiance fields (NeRF). Despite these attractive characteristics, segmenting and analyzing such intricate point clouds can be quite challenging. To effectively parse these complex geometric structures, a novel deep learning architecture named FieldCotSeg-Net is introduced. This model integrates an Anisotropy-aware Local Attention (ALA) module and a Hierarchical Feature Refinement Gate (HFRG) module to capture fine-grained features for precise point cloud segmentation. Experimental results demonstrate that the proposed model achieves outstanding performance on both the Huaxing3Dcot and Crops3D cotton datasets. On the Huaxing3Dcot dataset, the model yields a mean intersection over union (mIoU) of 75.7%, representing a 4.5% improvement over the baseline model. On the Crops3D cotton dataset, after retraining the model on this dataset, it attains an mIoU of 74.5%, showing substantial adaptability and effectiveness in organ segmentation. This technical system provides a scalable and robust solution for field cotton phenotyping analysis in breeding research and commercial production.

Plant phenotyping relevance

綿花の圃場表現型解析を目的に、UAVによる3D再構成と器官セグメンテーション手法を開発・評価しており、植物形態の取得が研究の中心である。

Code and data availability

The supplied blocks describe the self-constructed Huaxing3Dcot dataset (240 annotated cotton point clouds) and the FieldCotSeg-Net model, but no data availability statement, public repository, or authors' code/data URL appears in the provided text. The only URL present (Agisoft Metashape installer) is generic third-try

No evidence-backed public reproduction asset is currently recorded.