Abstract

High-resolution unmanned aerial vehicle (UAV) remote sensing images have super-high ground resolution. Although they provide complete and detailed surface observation data for various engineering applications, the extraction of information from complex and diverse surface scenes is challenging. Characterising surface targets with bright colours and different shapes using samples with fixed sizes and neural networks with fixed network structures at a single resolution is difficult. Therefore, a multi-resolution classification network called structure defined by sample characteristics (SDSC) network was designed in this study. After the SDSC network learned the samples using a multi-resolution strategy and the principle of maximum classification probability, the multi-resolution classification results were integrated into the final classification results to improve their credibility and accuracy. The new method has a better cognitive performance and noise resistance, as well as broad application potential, such that it is more suitable for high-spatial resolution UAV remote sensing images.

摘要

高分辨率无人机（UAV）遥感影像具有超高的地面分辨率。尽管它们为各种工程应用提供了完整而详细的地表观测数据，但从复杂多样的地表场景中提取信息具有挑战性。使用具有固定大小的样本和具有固定网络结构的神经网络以单一分辨率表征具有鲜艳颜色和不同形状的表面目标是困难的。因此，本研究设计了一种多分辨率分类网络，称为由样本特征定义的结构（SDSC）网络。SDSC网络利用多分辨率策略和最大分类概率原理学习样本后，将多分辨率分类结果整合到最终分类结果中，以提高其可信度和准确性。新方法具有更好的认知性能和抗噪声能力，具有广泛的应用潜力，更适用于高空间分辨率无人机遥感影像。