Research On Classification Method Of Remote Sensing Image Based On Full Convolutional Neural Network

Accurately obtaining information on ground features has a vital effect on all walks of life in human life.For example,the occurrence of large-scale soil salinization will not only lead to a sharp decline in land quality,but also reduce the production of cultivated land,water pollution,and even reduce biodiversity.Therefore,it is of great significance for humans to formulate policies and take measures to improve soil salinization by monitoring the distribution of saline-alkali land efficiently and accurately.However,with the improvement of the quality of remote sensing images,the shortcomings of large differences between ground objects and insufficient spectral information have gradually become apparent,which has brought huge challenges to the accurate and efficient classification of remote sensing images.At present,the two classification methods based on object-oriented classification and machine learning algorithm are the most common in high-resolution remote sensing image classification,but the two classification methods are time-consuming and laborious to be applied on a large scale,and both require manual selection and setting of parameters.Moreover,the model structure of machine learning algorithm is too shallow to extract deep features,so it is difficult to obtain satisfactory classification results.The deep learning classification method has sprung up in recent years,and has made great achievements in the field of image recognition.The deep learning model can automatically learn the deepest image features of the image,automatically set and constantly optimize the classification rules,thus carry out high-precision classification of images,which brings a more efficient and accurate classification method for remote sensing image classification.But,at present,deep learning models have not been applied to the classification of saline-alkali land,and there are fewer applications in a large scale.Therefore,this paper studies the full convolutional neural network FCN-8s model in deep learning.Take the classification of saline-alkali land in Baicheng City as an example,achieved the following results:1.A full convolutional neural network FCN-8s was built using Baicheng’s Gaofen-1 satellite PMS sensor 4-band data as the data set.The maximum likelihood classification map was used as pre-training data,and the visual interpretation map was used to fine-tune.The trained network obtained 86.20% accuracy on the test data set.2.Comparing the performance of the maximum likelihood method on the test data with the performance of this model,the classification accuracy is improved by 12.66%.The advantages and disadvantages of the two methods are analyzed.And the effectiveness of using the maximum likelihood classification graph as pre-training data is demonstrated.3.The NDVI value calculated from the data is added to the dataset as another band for training.The classification accuracy is 2% lower than that without NDVI.The reason for the accuracy reduction is analyzed from its principle and classification effect.4.All the images of Da’an in the Baicheng City are processed and input to the network,and all classification maps are processed into a map,which verifies the effectiveness of the method used in this paper and proves that this model can be effectively used in the study area.