当前位置: X-MOL 学术Forests › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
A Deep Learning Approach for Calamity Assessment Using Sentinel-2 Data
Forests ( IF 2.9 ) Pub Date : 2020-11-24 , DOI: 10.3390/f11121239
Daniel Scharvogel , Melanie Brandmeier , Manuel Weis

The number of severe storm events has increased in recent decades due to climate change. These storms are one of the main causes for timber loss in European forests and damaged areas are prone to further degradation by, for example, bark beetle infestations. Usually, manual mapping of damaged areas based on aerial photographs is conducted by forest departments. This is very time-consuming and therefore automatic detection of windthrows based on active and passive remote sensing data is an ongoing research topic. In this study we evaluated state-of-the-art Convolutional Neural Networks (CNNs) in combination with Geographic Information Systems (GIS) for calamity assessment. The study area is in in the northern part of Hesse (Germany) and was covered by twelve Sentinel-2 scenes from 2018. Labels of damaged areas from the Friedericke storm (18 January 2018) were provided by HessenForst. We conducted several experiments based on a custom U-Net setup to derive the optimal architecture and input data as well as to assess the transferability of the model. Results highlight the possibility to detect damaged forest areas using Sentinel-2 data. Using a binary classification, accuracies of more than 92% were achieved with an Intersection over Union (IoU) score of 46.6%. The proposed workflow was integrated into ArcGIS and is suitable for fast detection of damaged areas directly after a storm and for disaster management but is limited by the deca-meter spatial resolution of the Sentinel-2 data.

中文翻译:

使用Sentinel-2数据进行灾难评估的深度学习方法

由于气候变化,近几十年来严重风暴事件的数量有所增加。这些风暴是造成欧洲森林木材流失的主要原因之一,受损地区易于因树皮甲虫侵扰而进一步退化。通常,森林部门根据航拍照片手动绘制受损区域的地图。这非常耗时,因此基于主动和被动遥感数据自动检测风向是一个持续的研究主题。在这项研究中,我们结合了地理信息系统(GIS)评估了最新的卷积神经网络(CNN),以进行灾害评估。研究区域位于德国黑森州的北部,自2018年以来被十二个Sentinel-2场景所覆盖。HessenForst提供了Friedericke风暴(2018年1月18日)受损区域的标签。我们基于自定义的U-Net设置进行了几次实验,以得出最佳的体系结构和输入数据,并评估模型的可传递性。结果突出显示了使用Sentinel-2数据检测受损林区的可能性。使用二元分类,联合交叉口(IoU)得分为46.6%,实现了92%以上的准确性。拟议的工作流已集成到ArcGIS中,适用于在暴风雨后立即快速检测损坏区域并进行灾难管理,但受到Sentinel-2数据的十米空间分辨率的限制。我们基于自定义的U-Net设置进行了几次实验,以得出最佳的体系结构和输入数据,并评估模型的可传递性。结果突出显示了使用Sentinel-2数据检测受损林区的可能性。使用二元分类,联合交叉口(IoU)得分为46.6%,实现了92%以上的准确性。拟议的工作流已集成到ArcGIS中,适用于在暴风雨后立即快速检测损坏区域并进行灾难管理,但受到Sentinel-2数据的十米空间分辨率的限制。我们基于自定义的U-Net设置进行了几次实验,以得出最佳的体系结构和输入数据,并评估模型的可传递性。结果突出显示了使用Sentinel-2数据检测受损林区的可能性。使用二元分类,联合交叉口(IoU)得分为46.6%,实现了92%以上的准确性。拟议的工作流程已集成到ArcGIS中,适用于在暴风雨后立即快速检测损坏的区域并进行灾难管理,但受到Sentinel-2数据的十米空间分辨率的限制。联盟交叉口(IoU)得分为46.6%,实现了92%以上的准确性。拟议的工作流程已集成到ArcGIS中,适用于在暴风雨后立即快速检测损坏的区域并进行灾难管理,但受到Sentinel-2数据的十米空间分辨率的限制。联盟交叉口(IoU)得分为46.6%,实现了92%以上的准确性。拟议的工作流已集成到ArcGIS中,适用于在暴风雨后立即快速检测损坏区域并进行灾难管理,但受到Sentinel-2数据的十米空间分辨率的限制。
更新日期:2020-11-25
down
wechat
bug