基于深度学习服务的遥感影像农作物分类系统设计

Design of remote sensing image crop classification system based on deep learning service

关键词:  深度学习  作物分类  C/S  PaaS  CNN

Key words:  deep learning  crop classification  C/S  PaaS  CNN

摘要:

目的 深度学习在图像分类方面效果显著，但对机器的硬件配置要求高，将深度学习的技术应用于作物分类的同时，降低客户端的IT成本，开发基于深度学习服务的遥感影像农作物分类系统。方法 系统采用C/S架构，服务器端部署Caffe的深度学习框架，通过PaaS提供计算服务，统一处理客户端模型训练、影像分类等任务；客户端提供用户界面，负责数据输入、结果解析和可视化操作；客户端和服务器之间通过异步RPC实现网络通信，利用FTP进行数据的上传和下载；考虑到农作物分类在实际应用中的时效性，系统还提供了一个简化的Alexnet模型，在保证分类精度的前提下加速了模型的收敛速度。结果 通过对采集的428144个农作物样本数据集的应用表明，服务器训练的时间比单机训练缩短了近3倍，该系统不仅能够快速地完成深度学习模型训练的任务，还能实时准确得到遥感影像的农作物分类结果。结论 该研究进一步推动了深度学习在农业分类方面的应用，同时为遥感技术在农业应用中的发展、农作物面积统计工作和对农业资源进行优化配置提供了重要的科学指导。

Abstract:

Purpose With the rapid development of remote sensing technology，remote sensing has become an important means of obtaining farmland information in the agricultural field.How to quickly and accurately classify large-scale remote sensing images is a hot spot in the agricultural field.The popular deep learning is effective in image classification，but it requires high hardware configuration for the machine.Method Therefore，while applying deep learning techniques to crop classification，in order to reduce the IT cost of the client，this paper develops a crop classification system of remote sensing image based on deep learning services.The system with client/server architecture sets up a deep learning server.The server deploys Caffe’s deep learning framework，provides computing services through PaaS，and uniformly handles model training and image classification tasks of the client.The client is responsible for data input，result analysis and visualization operations，and provides user interface.The network communication between the client and the server is performed through asynchronous RPC，and data is transmitted by FTP.Considering the timeliness of crop classification in practical application，the system provides a simplified Alexnet network structure which reduces the number of network layers，and the convergence speed of the model is accelerated on the premise of ensuring accuracy.Result The application of the collected 428，144 crop samples shows that the training time of the server is nearly three times shorter than that of the stand-alone training.The system not only can quickly complete the task of deep learning model training，but also obtain the crop classification results of remote sensing images in real time and accurately.Conclusion The research further promotes the application of deep learning in agricultural classification，and provides important scientific guidance for the development of remote sensing technology in agricultural applications，crop area statistics and optimal allocation of agricultural resources.