基于深度学习的卫星影像综合类图斑信息提取探讨

Research on Composed Spot Information Extraction Based on Deep Learning

本文从重庆市年度批准未建设重点专项监测的实际需求出发,建立了图斑分类标准,构建了批准未建设用地的图斑样本库。以语义分割模型为基础,设计地物要素提取模型,并利用深度学习技术完成模型训练,实现了批准未建设用地综合类图斑的智能解译。试验结果表明,深度学习在综合类图斑信息提取方面有较好的应用前景,但样本库、训练底图和测试底图对测试结果的影响较大,因此提出了通过已有数据扩充样本库,针对不同数据源,开展模型训练,并增加后处理过程的新思路,以期提升模型预测效果,为进一步研究提供参考。

According to the actual needs of Chongqing's annual approval of unconstructed key special monitoring,the classification standard of map spots is established,and the map spot sample database of unconstructed land is constructed.Based on the semantic segmentation model,the feature element extraction model is designed.It uses deep learning technology to complete model training,and realizes the intelligent interpretation of approved non-construction land comprehensive map spots.The experimental results show that deep learning has a good application prospect in the comprehensive pattern spot information extraction,but the sample library,training map and test map have a greater impact on the test results.Therefore,a new idea is put forward to expand the sample base through existing data,carry out model training for different data sources and add post-processing process in order to improve the model prediction effect and provide reference for further research.