【摘要】 深度学习模型被广泛应用于农作物检测和识别领域，其优势在于通过构建不同的功能感知层而优化模型，能够自动提取输入数据的特征，实现端到端地学习。但是该模型中未知的数据处理过程导致模型缺乏可解释性，成为阻碍深度学习应用的主要因素。为克服深度学习模型可解释性不足的缺陷，研究者提出了基于类激活映射的可解释性方法。概述了类激活映射算法Grad-CAM在农作物病害分类和检测、农作物虫害检测识别、农作物品种分类、目标农作物检测以及其他应用上的研究进展，说明了类激活映射算法具有能够可视化任意结构卷积神经网络的优点，分析了类激活映射算法存在解释精细度不高、梯度不稳定、缺乏评估标准以及应用背景单一的不足，提出了构建既具有高准确率又具有可解释性的模型、构建新型解释算法、建立可解释性算法统一的评估标准和保证可解释性算法正确性的发展趋势。更多还原

【Abstract】 Deep learning models are widely used in the field of crop detection and recognition. Their advantage lies in optimizing the model by constructing different functional perception layers, which can automatically extract features from input data and achieve end-to-end learning. However, the unknown data processing process in this model leads to a lack of interpretability, which becomes the main obstacle to the application of deep learning. To overcome the shortcomings of insufficient interpretability in deep learning models, researchers have proposed an interpretability method based on class activation mapping. This article summarizes the research progress of the class activation mapping algorithm Grad-CAM in crop disease classification and detection, crop pest detection and recognition, crop variety classification, target crop detection, and other applications. It explains the advantages of the class activation mapping algorithm in visualizing convolutional neural networks with arbitrary structures, and analyzes the shortcomings of class activation mapping algorithms such as low interpretation precision, unstable gradients, lack of evaluation standards, and single application background. It proposes the development trend of building models with high accuracy and interpretability, construction of new interpretive algorithms, establishing unified evaluation standards for interpretable algorithms, and ensuring the correctness of interpretable algorithms.更多还原