Assessment of Three Automated Identification Methods for Ground Object Based on UAV Imagery
Assessment of Three Automated Identification Methods for Ground Object Based on UAV Imagery
复制标题
DOI：
10.3390/su142114603
复制
发表时间：
2022-11
期刊：
影响因子：
3.9
通讯作者：
Ke Zhang;Sarvesh Maskey;H. Okazawa;Kiichiro Hayashi;Tamano Hayashi;Ayako Sekiyama;S. Shimada
中科院分区：
文献类型：
--
作者：
Ke Zhang;Sarvesh Maskey;H. Okazawa;Kiichiro Hayashi;Tamano Hayashi;Ayako Sekiyama;S. Shimada
查明和监测实地的各种资源或废物对于综合资源管理十分重要。无人机（UAV）以其高分辨率和高机动性，成为精确、高效地监测地面目标的最佳工具。然而，以前的研究主要集中在土地利用和农艺分类方法的应用，很少有研究比较不同的分类方法使用无人机图像。将分类方法应用于地面目标识别，充分利用无人机的高分辨率是十分必要的。本研究比较了三种分类方法：A. NDVI阈值，B。基于RGB图像的机器学习，以及C.基于对象的图像分析（OBIA）方法A是耗时最少的，并且可以以高准确度（用户的准确度> 0.80）识别植被和土壤，但是在分类死亡植被、塑料和金属方面表现不佳（用户的准确度< 0.50）。方法B和C都是耗时耗力的，但在分离植被、土壤、塑料和金属方面具有非常高的准确性（用户对所有类别的准确性≥ 0.70）。方法B在识别具有明亮颜色的对象方面表现出良好的性能，而方法C在分离具有相似视觉外观的对象方面表现出很高的能力。从科学的角度验证了现有分类方法在识别小于1 m的小地物上的可行性，并讨论了3种方法精度不同的原因。实践上，这些结果有助于不同领域的用户选择适合他们目标的方法，从而可以通过组合不同的方法来同时监测不同的废物或多种资源，这有助于改进综合资源管理系统。
Identification and monitoring of diverse resources or wastes on the ground is important for integrated resource management. The unmanned aerial vehicle (UAV), with its high resolution and facility, is the optimal tool for monitoring ground objects accurately and efficiently. However, previous studies have focused on applying classification methodology on land use and agronomy, and few studies have compared different classification methods using UAV imagery. It is necessary to fully utilize the high resolution of UAV by applying the classification methodology to ground object identification. This study compared three classification methods: A. NDVI threshold, B. RGB image-based machine learning, and C. object-based image analysis (OBIA). Method A was the least time-consuming and could identify vegetation and soil with high accuracy (user’s accuracy > 0.80), but had poor performance at classifying dead vegetation, plastic, and metal (user’s accuracy < 0.50). Both Methods B and C were time- and labor-consuming, but had very high accuracy in separating vegetation, soil, plastic, and metal (user’s accuracy ≥ 0.70 for all classes). Method B showed a good performance in identifying objects with bright colors, whereas Method C showed a high ability in separating objects with similar visual appearances. Scientifically, this study has verified the possibility of using the existing classification methods on identifying small ground objects with a size of less than 1 m, and has discussed the reasons for the different accuracy of the three methods. Practically, these results help users from different fields to choose an appropriate method that suits their target, so that different wastes or multiple resources can be monitored at the same time by combining different methods, which contributes to an improved integrated resource management system.