A novel three-dimensional canopy photosynthesis model for tea plant: integrating point cloud deep learning and ray tracing to optimize photosynthetic efficiency

Abstract

The tea plant (Camellia sinensis) is economically and nutritionally important because of its bioactive compounds. Photosynthesis directly affects tea's growth and productivity, requiring a detailed study of its relationship with cultivation outcomes. We developed a novel computational pipeline for constructing three-dimensional (3D) canopy photosynthesis models of tea plant, leveraging multi-view stereo 3D reconstruction. The ISBNet architecture was optimized for precise leaf–stem segmentation from point cloud data, achieving 0.897 average precision (AP) for leaves and 0.793 AP for stems. We then created a plant leaf morphology-adapted meshing algorithm optimized for plant leaf morphology, achieving an average mesh reduction of approximately 96% while maintaining morphological fidelity compared with conventional meshing methods. We generated multiple tea plant canopies representing distinct planting patterns, and used a ray tracing algorithm to simulate the spatiotemporal distribution of light within these structures. Canopy photosynthesis simulation revealed significant cultivar-specific differences, with 'Yuehuang 1' exhibiting the highest photosynthetic activity. Dense planting (10 cm spacing) significantly enhanced canopy photosynthetic rates compared with wider spacing (20 cm), and a strong linear correlation (r = 0.99) was identified between total leaf area and daily canopy photosynthetic rate across cultivars. This work establishes a methodological foundation for precision agriculture optimization in perennial crops, providing quantitative guidance for maximizing tea plantations' productivity through optimal cultivar selection and spatial configuration.

Plant phenotyping relevance

茶樹キャノピーの3D再構築、葉・茎セグメンテーション、形態適応メッシュ化、光線追跡による光合成推定を統合した方法開発が中心であり、植物形態・光合成状態の定量化に直接つながる。

Code and data availability

The paper's tea plant point cloud datasets, annotated training data, ISBNet model, and analysis pipeline are not publicly deposited; the authors state data are available from the corresponding author upon reasonable request.

No evidence-backed public reproduction asset is currently recorded.