Morphological Recognition And Classification Of Urban High-Rise Buildings In Satellite Image

Automatic analysis of urban high-rise building block morphology in satellite images plays an important role in the automation of urban planning and research.The morphology of urban high-rise building generally includes residence and office.The main research of this subject includes two aspects,one is the high-rise building blocks,namely the automatic detection and classification of high-rise building communities,and the other is the prediction of the height parameters of high-rise buildings.Since there are few studies done,at present,all the data of this subject are marked by myself and reviewed by experts.Previous detection of high-rise buildings in optical satellite imagery has also relied on artificial properties.This topic assumes that tall building communities can be detected as a target,proposing a tall building community detection method based on neural network.The method is divided into three steps: first,the high-rise building community and the building shadow are detected respectively,and then use the method of target filtering to obtain the final detection results of the high-rise building community.This method requires only optical satellite images without combining other data,reducing additional cost.Experimental results show that the method proposed in this topic is Precision 0.96,Recall 0.79 and F1 Score 0.87.This method also confirms that high-rise building communities can be detected as a target in optical satellite images,and can be distinguished from other low-rise buildings.The above methods have limited detection effect on satellite images acquired in different cities and at different times.Aiming at this problem,this subject proposes an automatic detection method of high-rise building community based on multi-feature fusion based on neural network.The method firstly extracts the features of high-rise building communities through two different neural networks,and then uses the other two neural networks to fuse the extracted features to obtain preliminary detection results,and finally obtains the final detection results through a simple screening method.The feature of this method is that the extraction and combination of features are automatic,which is different from the previous method requiring manual intervention,which is conducive to the automation of analysis.By evaluating the effect of this method in detecting high-rise building communities in 15 high-resolution optical satellite images of different cities,the average Precision is 0.79,Recall is 0.77,and F1 Score is 0.78.The method also confirms that high-rise building communities have similar characteristics in satellite images of different cities.Previous studies have relied on manual calculations of fixed formulas for obtaining the height parameter.This topic assumes that the height of tall buildings can be predicted by automatically extracting features from optical satellite images.Therefore,this subject proposes a multi-task neural network method combining object detection task and regression prediction to predict the height parameters of high-rise buildings in optical satellite images.The method extracts features for a single high-rise building,detects a single building area in the image,and predicts height parameters through a fully connected layer.Through evaluation,the method can accurately detect the building area and more accurately predict the height value.The two research directions of this topic provide some theoretical reference and practical experimental data support for the automation of morphological analysis of high-rise building blocks.Increasing the amount of data and improving the neural network structure can improve the automatic detection accuracy.