Crop diseases pose as a major threat to global food security. Minimizing disease-induced damage during crop growth and optimizing crop yields are vital for agricultural sustainability. Therefore, advanced disease detection and prevention of such diseases are crucial and the detection must be prompt and efficient as it is essential for the implementation of appropriate control measures. In this work, a parallel deep learning framework based on deep feature fusion is developed to precisely identify the severity of crop diseases. The framework utilizes ResNet50 and Xception as separate branches for feature extraction. Convolutional layer weights are initialized through transfer learning techniques employing models pre-trained on the ImageNet dataset. A fine-tuning strategy is employed for the optimization of convolutional layers and the design of the top layer. This framework achieves an accuracy of 88.58% on the AI Challenger 2018 dataset, marking an enhancement of 2.8% and 13% over other influential deep learning models, and it also outperforms some recent works. Likewise, the framework exhibits a recognition accuracy of 99.53% on the PlantVillage dataset. Moreover, an Android-based application is developed to diagnose the severity of crop diseases in real-time. The diagnostic system swiftly procures results and provides control recommendations through the capture and upload of images to the platform. The advanced severity detection system reduces the expertise required from users, facilitating precise prevention and control measures whilst focusing on accessibility. This work aims to provide innovative approaches and solutions for disease detection in the agricultural field, utilizing artificial intelligence to enhance agricultural informatization and creating more sustainable farming methods.

中文翻译：

---

一种利用深度迁移学习估计农作物病害严重程度的新型移动诊断系统


农作物病害对全球粮食安全构成重大威胁。最大限度地减少作物生长过程中疾病造成的损害并优化作物产量对于农业可持续发展至关重要。因此，先进的疾病检测和预防至关重要，检测必须及时有效，因为这对于实施适当的控制措施至关重要。在这项工作中，开发了基于深度特征融合的并行深度学习框架，以精确识别农作物病害的严重程度。该框架利用 ResNet50 和 Xception 作为特征提取的单独分支。卷积层权重通过采用在 ImageNet 数据集上预训练的模型的迁移学习技术进行初始化。采用微调策略来优化卷积层和顶层的设计。该框架在 AI Challenger 2018 数据集上实现了 88.58% 的准确率，比其他有影响力的深度学习模型分别提高了 2.8% 和 13%，并且也优于最近的一些工作。同样，该框架在 PlantVillage 数据集上的识别准确率达到 99.53%。此外，还开发了一款基于Android的应用程序来实时诊断农作物病害的严重程度。诊断系统通过捕获图像并将其上传到平台来快速获取结果并提供控制建议。先进的严重性检测系统减少了用户所需的专业知识，促进精确的预防和控制措施，同时注重可访问性。 这项工作旨在为农业领域的疾病检测提供创新的方法和解决方案，利用人工智能增强农业信息化，创造更可持续的耕作方法。