ABSTRACT

In this paper, a different approach based on convolutional neural networks (CNNs) is proposed to generate digital surface model (DSM) from a single high-resolution satellite image. In this regard, an approach based on a deep convolutional neural network was designed. The proposed CNN has an encoder-decoder structure to extract multi-scale features in the encoding part and estimate the height values by up-sampling the extracted abstract features. Then, a filtering approach based on morphological operators is proposed to extract the non-ground pixels from each estimated height image. The final digital surface mode Shuttle Radar Topography Mission (SRTM) is obtained by integrating the SRTM elevation model and extracted non-ground objects. Evaluating the estimated height images indicated 0.219, 0.865, and 2.912 m on average log10 error, relative error, and root mean square error (RMSE), respectively. In addition, investigating the final integrated DSM indicated 4.625 m on average for RMSE, demonstrating a promising performance of the proposed approach.

摘要

在本文中，提出了一种基于卷积神经网络（CNN）的不同方法，以从单个高分辨率卫星图像生成数字表面模型（DSM）。在这方面，设计了一种基于深度卷积神经网络的方法。所提出的CNN具有编码器-解码器结构，以在编码部分中提取多尺度特征，并通过对所提取的抽象特征进行上采样来估计高度值。然后，提出了一种基于形态学算子的滤波方法，从每个估计的高度图像中提取非地面像素。通过整合SRTM高程模型和提取的非地面物体，可以获得最终的数字水面模式航天飞机雷达地形任务（SRTM）。对估计的高度图像进行评估后，平均log10误差，相对误差为0.219、0.865和2.912 m，和均方根误差（RMSE）。此外，对最终的综合DSM进行的调查显示，RMSE平均为4.625 m，表明了所提出方法的良好前景。