基于ResNet50的端到端深度学习模型在玉米干旱胁迫多光谱监测中的应用OA

With the increasing severity of drought threats due to global climate change,developing accurate and efficient crop stress monitoring methods is crucial for ensuring food security.The performance of end-to-end deep learning and traditional machine learning models in classifying drought stress levels in maize was evaluated and compared.Six-channel multispectral images of 8 maize inbred lines under 5 different water treatments were collected using a high-throughput phenotyping platform.Based on this data,an end-to-end deep learning model with a ResNet50 architecture and a random forest machine learning model built on 48 manually extracted phenotypic features were constructed and compared.The results demonstrated that the ResNet50 model achieved superior classification performance,with a macro-average F1-score of 0.887,significantly outperforming other benchmark models.Interpretability analysis via channel masking further revealed that the exceptional performance of ResNet50 primarily stemmed from its ability to autonomously learn complex stress-related patterns from the red and near-infrared spectral bands,which extended beyond the limits of human visual perception.In contrast,the random forest model relied mainly on morphological parameters and the chlorophyll carotenoid index.To sum up,the end-to-end deep learning models possessed distinct advantages over traditional feature engineering methods for crop stress grading using high-dimensional multispectral data,providing reliable technical pathway for the development of intelligent precision irrigation and crop health monitoring systems.