Design and Application of an Intelligent Plant Disease and Pest Recognition System for Landscape Architecture Based on Deep Learning

Abstract

Manual inspection of landscape pathology and textile fiber defects suffers from inherent subjective bias and suboptimal throughput. To bypass these bottlenecks, we propose the Ghost-Convolution Enlightened Vision Transformer (GeT). We constructed a novel hybrid neural network architecture, the Ghost-Convolution Enlightened Vision Transformer (GeT), which synergistically integrates the lightweight local feature extraction proficiency of Convolutional Neural Networks (CNN) with the global semantic modeling capabilities of Vision Transformers (ViT). Utilizing a newly established standard dataset (GLDP15k) comprising 15, 000 heterogeneous field images, the model was subjected to rigorous hyperparameter optimization and ablation studies. Optimization on the GLDP15k dataset yielded a peak accuracy of 96.8% across 12 target classes, maintaining a Kappa-coefficient of 0.941. Constrained to 1.16 M parameters, the architecture executes at 5.5 ms per image (180 FPS) on edge hardware. A 6-month application at Yuexiu Park demonstrated a 3.2-fold improvement in detection efficiency and a 35% reduction in pesticide usage compared to manual inspections. This study not only elucidates the interpretability of hybrid attention mechanisms in phytopathology but also adapts these vision-based paradigms to the detection of microscopic anomalies in textile weaving patterns, providing a scalable and computationally efficient solution for both precision plant protection and industrial fabric defect inspection.

Plant phenotyping relevance

植物病害を画像から認識する深層学習モデルを開発し、専用データセットで検証・応用しており、植物の病害状態の取得方法が中心的な研究貢献である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.