Scene Classification Of High-Resolution Remotely Sensed Images Based On DCNN

With advances in geological research,economic growth,rapid social development and the many factors accompanying development,remote sensing technology for earth observation is becoming more important,and remotely sensed data has entered the era of high-resolution imaging.Compared with middle and low resolution remotely sensed images,high spatial resolution remotely sensed images can provide more shape,texture and spatial information,and have broad application prospects in fine agriculture,cadastral survey,urban planning and many other fields.However,traditional remotely sensed image classification methods cannot meet the needs of high-level remotely sensed image content interpretation.Therefore,high resolution remotely sensed image classification for scenes is a hot research topic at present.Gf-2satellite,as China’s first independently developed sub-meter-level remote sensing satellite with spatial resolution,can acquire full-color images with spatial resolution of 0.8m and multi-spectral images with spatial resolution of 3.2m.It has rich spatial information and broad application prospects.In this paper,the depth of the high resolution remotely sensed images features extracted based on the deep convolution neural network（DCNN）structure of the residual learning network（ResNet）,and the semantic features of UC Merced Land Use（UCM）data set were constructed together with the low-level features of high-resolution remotely sensed images to train two kinds of classifiers,including support vector machine（SVM）and K-Nearest neighbor（KNN）algorithm,and compare the classification results to get high resolution remotely sensed images scene classification model.Then,we set up 7 GF-2 scene data sets,and use the sample migration method in migration study theory,to scene classification of GF-2 data set.Finally,scene classification of GF-2 images was realized to provide the research foundation for the application of high resolution remotely sensed image in more fields.It mainly includes the following research results:1)Extraction based on ResNet network and the depth of the high resolution remotely sensed image feature extracting the underlying characteristics of high resolution remotely sensed image color moment and gray level co-occurrence matrix characteristics,through the different characteristics of the combination,constitute four express the combination of high resolution remotely sensed image scene semantic features,including Feature_R,Feature_RC,Feature_RG,Feature_RCG,used for high resolution remotely sensed image classification.2)Based on the above four combined features,eight classification models of SVM and KNN classifiers were respectively trained to form scene classification models for UCM dataset.After precision evaluation,comparative analysis showed that under the SVM classifier,scene classification accuracy by using Feature_RC and Feature_RCGCG reached 93.81%,which is better than other classification models.Therefore,the Feature_RC-SVM classification model is the optimal classification model of high-resolution remotely sensed image scenes in this paper.3)According to migration learning theory,seven scene images of UCM data set including agriculture,dense residential area,forest,intersection,overpass,parking lot and river are selected as migration samples,and Feature_RC-SVM classification model is adopted to realize scene image classification of GF-2 dataset with small sample size,and the classification accuracy is up to 95.71%.Finally,scene boundary of the selected research area was combined to conduct scene level classification of GF-2images,and the classification results were consistent with the reality.It can provide research basis for further target detection and recognition of GF-2 images.