航拍死角、匹配偏差、点位不足、云雾遮盖等易使数字高程模型(DEM)存在数据异常.鉴于光学立体摄影测量获得的ASTER GDEM受天气因素影响大,雷达测量得到的SRTM DEM受地形起伏因素影响大,为提高AS-TER GDEM数据的质量,该文构建一种多源多尺度残差连接门控卷积生成对抗网络(Multi-Source and Multi-Scale Residual-Connected Gated Convolutional Generative Adversarial Network,MSSRGC-GAN),利用 SRTM DEM 数据辅助重建ASTER GDEM中异常数据,并以实验区范围内多组不同地貌的ASTER GDEM样本数据重建为例,对模型进行定量评价.结果显示:重建数据RRMSE小于0.06,R2大于0.9,PSNR大于60,SSIM在0.999 5以上,优于反距离插值法和SRTM镶嵌法等传统方法以及无门控卷积模型和无空洞卷积模型等深度学习方法.
Integration of Generative Adversarial Networks and SRTM for Reconstruction of ASTER GDEM Data
Digital elevation models(DEMs)frequently exhibit data anomalies attributable to various factors,including occlusions present in aerial photography,discrepancies from matching errors,sparse point coverage,and interference from cloud cover.The advanced spaceborne thermal emission and reflection radiometer global digital elevation model(ASTER GDEM),derived through optical stereophotogrammetry,is significantly affected by weather-related variables.Conversely,the shuttle radar topog-raphy mission(SRTM)digital elevation model(DEM),which is generated from radar-based surveys,is predominantly influ-enced by the topographic relief of the terrain.In order to improve the quality of ASTER GDEM data,a multi-source and multi-scale residual-connected gated convolutional generative adversarial network(MSSRGC-GAN)is constructed.This model levera-ges SRTM DEM data to assist in reconstructing the abnormal data in ASTER GDEM.In order to quantitatively assess the mod-el outcomes,multiple sets of ASTER GDEM sample data with different terrains within the validation area were reconstructed.The reconstructed data yielded RRMSE below 0.06,R-above 0.9,PSNR above 60,and SSIM above 0.999 5.This performance surpasses that of conventional techniques such as inverse distance interpolation and SRTM mosaicking,as well as deep learning methods including the ungated convolutional model and the non-dilated convolutional model.
万方数据
维普
