Research On Small Farmlands Recognition Method Based On Fully Convolutional Networks

Recently,with the progress of science and technology,the spatial resolution of remotesensing image has been improved constantly.High-resolution remote sensing image not only has rich spatial and texture features,but also contains a large number of scene semantic information,which results in the greatly increasing difficulty of small items identifying in the image.Traditional identification methods based on spectral statistical characteristic,such as Maximum Likelihood Estimate(MLE)and K-Nearest Neighbor(KNN)only use the spectral information of the image,not fully utilizing the rich detail information in the image.Therefore,the requirements of the extraction of tiny farmlands cannot be satisfied.In view of the above problem,the research takes GF-1 high-resolution remote sensing images as the data source.At the same time,the research applies the fully convolutional network algorithm which is very popular in the deep learning to the recognition of tiny farmlands in highresolution remote sensing image.In general,our contributions in this paper mainly consist of the following two parts:Aiming at the problems of less effective information and less expression of small farmlands in image,and many difficulties in traditional object recognition methods,a farmlands enhancement algorithm for training sample images based on the denoising sparse automatic encoder(SDA)is proposed in this paper.Multiple SDAs are stacked to obtain a kind of automatic encoder for stacking and noise reduction.Further,the L-BFGS algorithm is introduced to train and excavate deeply the most important high-level abstract features in the input images.The features learned automatically,as the primary filter of the original signal,are added to the original features,so as to attain the enhanced sample images.In this way,the information quantity of small farmlands is enhanced,which provides high-quality data support for subsequent training more accurate small farmlands recognition model.Based on the fully convolutional networks(FCN)algorithm,a small farmlands recognition model for high resolution remote sensing images is established through combining with the convolution neural network model VGG-19 under the TensorFlow framework.According to the VGG-19 model,we build the network structure,and the upper sampling layers are added at the end of the model.Further,we employ the model parameters in VGG-19 to initialize the network structure parameters built in our research.Then,aiming at the problem that the boundary of small farmlands in recognition results is blurred and the recognition accuracy is low,on the one hand,we combine the advantages of the two functions of Sigmoid and ReLU to act as the activation function in the model by stacking them.On the other hand,the probability values outputted by Softmax classification level in Fully Convolutional Networks are combined with the naive Bayes discriminant theory.The probability values calculated by the naive Bayes discriminant theory are combined with the Softmax layer output values of the corresponding regions in order to obtain the ultimate probability values of the data points in each category.Experiments have shown that the trained model can achieve better recognition results and effectively improve the accuracy of small-block farmland recognition.