基于深度学习的农业大棚遥感提取方法研究

Research on Remote Sensing Extraction Method of Agricultural Greenhouse Based on Deep Learning

及时、准确地监测农业大棚的位置以及空间分布,是开展耕地“非农化”整治的工作重点。遥感技术因其时效性强、覆盖范围广等优势,已逐渐成为大棚监测的主要手段。为了提高农业大棚的遥感识别精度,缓解深度学习方法对于大数据量的需求,提出了一种基于残差神经网络(Residual Neural Network,ResNet)和迁移学习的农业大棚自动识别方法,能有效区分大棚及其混淆地物(林地、耕地、建设用地和水体)。将经过ImageNet图像数据集预训练得到的网络权重加载到ResNet网络模型中,利用少量经过数据扩增的训练样本对模型参数进行微调,将模型迁移至遥感大棚识别问题。采用Sentinel-2中分辨率卫星影像对所提方法进行了验证,实验结果证明,经过迁移学习的ResNet模型的大棚识别率达到了95.5%,与未经过迁移学习的模型相比精度提升了1.6%,能有效提高小样本条件下大棚识别的准确性。

Timely and accurate monitoring of the location and spatial distribution of agricultural greenhouses is the focus of the“non-agricultural”rectification of cultivated land.Remote sensing technology has gradually become the main means of greenhouse monitoring because of its strong timeliness and wide coverage.In order to improve the remote sensing recognition accuracy of agricultural greenhouse and alleviate the demand of deep learning method for large amount of data,an automatic recognition method of agricultural greenhouse based on residual neural network(ResNet)and transfer learning is proposed,which can effectively distinguish between greenhouse and other confusing ground objects(forest land,cultivated land,construction land and water body).The network weight obtained from the pre-training of ImageNet image data set is loaded into the ResNet network model,and then the model parameters are finely tuned through a small number of data amplified training samples to transfer the model to the remote sensing greenhouse recognition problem.Sentinel-2 medium resolution satellite images are used to verify the above methods.The experimental results show that the greenhouse recognition rate of ResNet model with transfer learning is 95.5%,and the accuracy is improved by 1.6%as compared with the model without transfer learning.The results show that the algorithm can effectively improve the accuracy of greenhouse recognition under the condition of small samples.