Precision Detection of Salt Stress in Soybean Seedlings Based on Deep Learning and Chlorophyll Fluorescence Imaging.

Abstract

Soil salinization poses a critical challenge to global food security, impacting plant growth, development, and crop yield. This study investigates the efficacy of deep learning techniques alongside chlorophyll fluorescence (ChlF) imaging technology for discerning varying levels of salt stress in soybean seedlings. Traditional methods for stress identification in plants are often laborious and time-intensive, prompting the exploration of more efficient approaches. A total of six classic convolutional neural network (CNN) models-AlexNet, GoogLeNet, ResNet50, ShuffleNet, SqueezeNet, and MobileNetv2-are evaluated for salt stress recognition based on three types of ChlF images. Results indicate that ResNet50 outperforms other models in classifying salt stress levels across three types of ChlF images. Furthermore, feature fusion after extracting three types of ChlF image features in the average pooling layer of ResNet50 significantly enhanced classification accuracy, achieving the highest accuracy of 98.61% in particular when fusing features from three types of ChlF images. UMAP dimensionality reduction analysis confirms the discriminative power of fused features in distinguishing salt stress levels. These findings underscore the efficacy of deep learning and ChlF imaging technologies in elucidating plant responses to salt stress, offering insights for precision agriculture and crop management. Overall, this study demonstrates the potential of integrating deep learning with ChlF imaging for precise and efficient crop stress detection, offering a robust tool for advancing precision agriculture. The findings contribute to enhancing agricultural sustainability and addressing global food security challenges by enabling more effective crop stress management.

Plant phenotyping relevance

深層学習とクロロフィル蛍光画像から、ダイズ幼植物の塩ストレス状態を分類する画像ベースの表現型取得・解析手法が研究の中心である。

Code and data availability

The paper's ChlF image dataset (421 original images), SPAD measurements, and analysis code are not publicly deposited; the Data Availability Statement says data are available only on request. No public URL or repository is provided.

No evidence-backed public reproduction asset is currently recorded.