Cost-Effective Smartphone-Based Computer Vision Pipeline for Actinidia Phenological Stage Mapping

Abstract

Abstract Dioecious crops face significant pollination challenges due to the asynchrony in flowering between male and female plants. This asynchrony varies spatially across orchards, requiring targeted interventions in zones where synchrony is lacking. Assisted pollination addresses this deficiency, albeit at a substantial operational cost that could be optimised through spatial phenological mapping. Manual assessment proves economically infeasible at commercial scales, while existing computer vision systems are unable to classify phenological stages and integrate geospatial information. This study presents the Mobile Phenological Mapping (MPM) Framework, which integrates GNSS-synchronised smartphone video with automated phenological detection to generate plant-level phenological distribution maps validated in commercial kiwifruit (Actinidia chinensis) orchards. This modular framework comprises (i) training and validation data acquisition, (ii) model optimisation, (iii) operational pipeline, and (iv) performance evaluation. MPM employs hierarchical deep learning across three stages: structure detection, gender classification, and phenological stage classification. Video frames are georeferenced through timestamp matching with GNSS metadata, enabling spatial phenological mapping. Operational validation across four commercial orchard zones demonstrated mean absolute percentage errors of 17.2% for structure detection and 20.1% for gender classification. The framework reduces monitoring time from 113 to 1.6 hours per hectare, decreasing labour costs from €2 060 (113 hours × 18.20 € per hour) to €29 (1.6 hours × 18.20 € per hour) per hectare based on the Portuguese hourly labour cost for minimum wage workers. When integrated with routine orchard operations, video acquisition incurs negligible additional cost. MPM provides growers with precision phenological maps for targeted pollination interventions. While validated in a kiwifruit orchard, the modular architecture can be adapted to other crops by replacing the training data.

Plant phenotyping relevance

スマートフォン動画とGNSSを用いて植物体レベルの性別・生育段階を自動検出し、フェノロジー分布を作成する手法と、その性能評価・商業園での検証が研究の中心である。

Code and data availability

植物フェノタイピング解析を再現する公開資産であることを、入力本文と直接リンクから確認できなかったため保留しました。

No evidence-backed public reproduction asset is currently recorded.