Multiple Object Extraction of Remote Sensing Images Based on Convolutional Neural Networks and Support Vector Machines

Tao-xing Feng,Wei-yu Wang,Xiao-meng Qian,Zhan-hong Zhang,Hui Liu,Ying Xing,Bin Yang

DOI: https://doi.org/10.12783/dtcse/cnai2018/24182

2018-01-01

DEStech Transactions on Computer Science and Engineering

Abstract:The data provided by remote sensing technology is characterized by wide coverage, high real-time performance and a supply of rich and objective information. Therefore, extracting geo-objects such as roads from remote sensing images plays important roles in many urban applications and used in a wide range of areas. However, in the past, such tasks generally performed manually, which is very costly and time-consuming. And many of existing automating attempts are designed for specific features and classifiers with limitation. This paper proposes a CNN-SVM-based road extraction system (a convolutional neural network with three support vector machines), which takes raw pixel values in aerial imagery as inputs and predicted three-channel label images (background-buildings-roads) as outputs. By training a single CNN efficiently, feature extractors and classifiers are automatically constructed and multiple kinds of objects are extracted simultaneously. Using SVM, the classification results can be optimized. At the same time, we improve our experimental performance by the Dropout method and predicting with spatial displacement. Finally, comparing our model with previous methods, we get the result that the precision of CNN-SVM increases 8.37%.