Abstract

Forest roads are the primary infrastructure facilities of forestry activities. Identifying the forest roads that can be used after a disaster is very important for management of disasters such as fire, flood, and landslide as well as for strengthening of the forest road infrastructure. In this context, forest road extraction from orthophoto images has become a hot research topic in the field of remote sensing image analysis. Deep learning methods stand out in many fields that require remote sensing and these methods lead to very successful results compared with traditional methods. Recently, deep learning methods are applied frequently in road extraction. The aim of the present study was forest road network extraction from high resolution orthophoto images based on deep learning. Four different deep learning models have been used in the study which are AlexNet, ResNet-50, InceptionResNet-V2 and U-Net. First, the images in the dataset were subject to re-processing after which the deep learning models were trained separately. Secondly, values of overall accuracy, precision, recall, Dice coefficient, intersection over union, and test time were calculated for these trained network models over the validity dataset. Finally, the acquired results were compared and forest road segmentation inferences were visualized thus putting forth the accuracy at which the deep learning models used can extraction the forest roads. The results show that ResNet-50 and InceptionResNet-V2 semantic segmentation models can be used accurately and efficiently for forest road extraction.

摘要

林道是林业活动的主要基础设施。确定灾后可使用的林道对于火灾、洪水、滑坡等灾害的管理以及加强林道基础设施具有重要意义。在此背景下，正射影像中的林道提取已成为遥感影像分析领域的研究热点。深度学习方法在许多需要遥感的领域中脱颖而出，与传统方法相比，这些方法取得了非常成功的结果。最近，深度学习方法在道路提取中得到了频繁的应用。本研究的目的是基于深度学习从高分辨率正射影像中提取森林道路网络。研究中使用了四种不同的深度学习模型，它们是 AlexNet、ResNet-50、InceptionResNet-V2 和 U-Net。首先，数据集中的图像经过重新处理，然后分别训练深度学习模型。其次，在有效性数据集上为这些训练有素的网络模型计算了总体准确度、精确度、召回率、Dice 系数、联合交集和测试时间的值。最后，对获得的结果进行比较，并对林道分割推断进行可视化，从而提出所使用的深度学习模型提取林道的准确性。结果表明，ResNet-50 和 InceptionResNet-V2 语义分割模型可以准确高效地用于林道提取。数据集中的图像经过重新处理，然后分别训练深度学习模型。其次，在有效性数据集上为这些训练有素的网络模型计算了总体准确度、精确度、召回率、Dice 系数、联合交集和测试时间的值。最后，对获得的结果进行比较，并对林道分割推断进行可视化，从而提出所使用的深度学习模型提取林道的准确性。结果表明，ResNet-50 和 InceptionResNet-V2 语义分割模型可以准确高效地用于林道提取。数据集中的图像经过重新处理，然后分别训练深度学习模型。其次，在有效性数据集上为这些训练有素的网络模型计算了总体准确度、精确度、召回率、Dice 系数、联合交集和测试时间的值。最后，对获得的结果进行比较，并对林道分割推断进行可视化，从而提出所使用的深度学习模型提取林道的准确性。结果表明，ResNet-50 和 InceptionResNet-V2 语义分割模型可以准确高效地用于林道提取。在有效性数据集上计算这些训练有素的网络模型的交集和测试时间。最后，对获得的结果进行比较，并对林道分割推断进行可视化，从而提出所使用的深度学习模型提取林道的准确性。结果表明，ResNet-50 和 InceptionResNet-V2 语义分割模型可以准确高效地用于林道提取。在有效性数据集上计算这些训练有素的网络模型的交集和测试时间。最后，对获得的结果进行比较，并对林道分割推断进行可视化，从而提出所使用的深度学习模型提取林道的准确性。结果表明，ResNet-50 和 InceptionResNet-V2 语义分割模型可以准确高效地用于林道提取。