Research On Cultivated Land Extraction Technology Of Remote Sensing Images Based On High Resolution Full Convolutional Network

Accurate and efficient extraction of cultivated land is the basis of agricultural production.It is of great significance for agricultural resource monitoring and national food security.With the development of remote sensing technology,extracting cultivated land through remote sensing imagery classification methods has become a very efficient method.For cultivated land in remote sensing images,there are large differences in the features such as the spectrum and texture,which makes it difficult to extract cultivated land.The traditional classification methods for remote sensing image use a shallow structure and requires manual participation in parameter selection and feature selection,which cannot effectively extract the features required for classification,resulting in poor extracting the same land-cover type which has different cover spectrums.This thesis expects to use the powerful feature extraction capabilities of deep learning methods to solve this problem,and explores the application of deep learning method transfer to the extraction of cultivated land.In transfer process,remote sensing images include not only the red,green,and blue band data commonly used in natural images,but also more band data such as nearinfrared.Exploring the effectiveness of each band for improving the results of cultivated land extraction is of great significance for making full use of rich information of remote sense images.Therefore,this thesis selects the bands of Landsat and Gaofen-2 satellite images,and constructs four datasets of different band combination.At the same time,this thesis selects the lightweight deep full convolution network U-Net,to avoid the problem that the pre-trained model obtained on the large natural image dataset is only applicable to three-band image data,and the method of re-training the U-Net model on the dataset with different band data is used for the extraction of cultivated land.In addition,due to the down-sampling operation in the U-Net network structure,the high-resolution details of images are easy to lose,leading to problems such as inaccurate details and smooth edges.Based on the full convolutional neural network UNet,we improve its skip connection and the input of its loss function,and propose a high-resolution U-Net algorithm(HRU-Net).Through experimental verification,HRU-Net and U-Net can extract the same landcover type correctly.At the same time,HRU-Net retains more high-resolution detail in the image than U-Net,improves the richness of the details of the cultivated land extraction results,and further improves the accuracy of cultivated land extraction.