Research On Island Township Information Management Platform Based On Multi-source Data

With the advancement of “Construction of 3D realistic geospatial scene of China”,the data cockpit in the form of charts and graphs is gradually less helpful in the daily management of towns and villages,as well as in assisting them to make decisions.At the same time,3D real scene management platforms with realistic scenes and good interactivity continue to emerge.In order to develop the application in a realistic scenario,a 3D model of the towns and villages needs to be reconstructed and the land types need to be automatically segmented and classified.However,the houses in towns and villages are tightly built,and the buildings block each other,there are localised deformations and distortions when directly use UAV tilt photography.This dissertation takes Ma’ao Town as an example to build a Multi-Source data Information Management Platform on island villages and island towns.Firstly,tilt photogrammetry is carried out by UAVs equipped with Li DAR cameras and full-frame cameras,supplemented by UAV surround photography.The collected multi-source data are fused to build a comprehensive,complete and highly accurate 3D real scene model of the town.Secondly,the model is segmented and extracted from the image based on the land type.However,due to the complexity of the feature scenes in the island towns and island villages,it is time-consuming and laborious to use traditional manual labelling.This dissertation thus introduces a semantic segmentation model for the accurate segmentation and classification of 3D real scene models of towns and villages.The model is optimised and improved on the basis of the Deep Labv3+ semantic segmentation model to make it more suitable for segmentation of land type diversity.The main features include: using the deep residual network Res Net+ as the backbone network to improve the training efficiency and performance of the network;adding a joint upsampling module to improve the network’s ability to recognise objects at different scales and the resolution of feature maps;adjusting the void space pyramid pooling module to reduce the combination of expansion rates to improve the recognition accuracy of the network;improving the decoder to better recover detail information and edge information to improve the accuracy of the network.The training and testing results of the dataset in this dissertation show that compared to the traditional model,the model in this dissertation has improved in terms of accuracy and average delivery ratio,and fully satisfies the demand for realistic 3D model segmentation of towns and villages to achieve land type classification.Finally,the construction of thematic applications for land resources and management review is targeted according to the needs analysis.Using the advantages of 3D real scene,Multi-Source data Information Management Platform on island villages and island towns provides visual data analysis and decision support for the government,and provides more high-quality and efficient services for related industries,thus promoting the economic development and industrial upgrading of towns and villages.