International Journal of Applied Earth Observations and Geoinformation (Feb 2023)
Using object-oriented coupled deep learning approach for typical object inspection of transmission channel
Abstract
Traditional methods to inspect transmission line passage are mostly through UAV/helicopters and manual inspection, which are limited in no man's area and large-scale inspection. Satellite remote sensing technology breaks through such limitations and makes it more possible to inspect transmission line passage on a large scale in the power industry. Therefore, adopting the WorldView-3 satellite image with high resolution as the data source, we propose a typical ground object inspection method based on the object-oriented method coupled with the deep learning method. Firstly, in view of the traditional object-oriented multicategory extraction phenomenon, we use multiscale segmentation technology to construct the typical feature classification rule set and optimize the segmentation threshold. Through prior knowledge and experiments, we explore the best inspection sequence of multicategory typical features. Secondly, we use a fully convolutional neural network to perform end-to-end inspections of typical features. Finally, we propose the combination of the two methods to make it act on an image by comparing the inspection results of each typical ground object category and analyzing the advantages and disadvantages of the above two methods. Based on the principle of optimal inspection sequence, the user flexibly selects target extraction methods and finally the typical ground object inspection results are output end-to-end. The research shows that the coupled method has the highest inspection accuracy. The overall inspection accuracy reaches 93.01%, which is 10.32% and 8.07% higher than that of the object-oriented method and the deep learning method. The Macro-F1 score is 0.8363, which is 0.124 and 0.1532 higher than that of the object-oriented method and the deep learning method. We also choose Lingwu City for a verification experiment. The result shows that the overall inspection accuracy for the new experimental area is 91.25%, and the Macro-F1 score is 0.8948, indicating that the algorithm has versatility.
