Exploring unsupervised domain adaptation methods for automated linear disturbance mapping

在加拿大北部广阔的北方森林和湿地地区，道路、地震勘探、管道和能源传输走廊等干扰是导致林地北美驯鹿（Rangifer tarandus）——北方森林种群减少的主要原因。因此，对这些“线性扰动”的深刻理解已成为加拿大研究和森林管理的重点。支持管理线性扰动的理想工具是一种自动生成具有成本效益的地图的方法，该地图可以准确地识别这种形式的森林栖息地破碎化。因此，该项目的重点是开发一种使用深度学习方法从多光谱卫星图像自动生成地图的方法。具体来说，用于对卫星图像中的每个像素进行分类的卷积神经网络将用于从Sentinel-2数据生成线性扰动图。需要解决的主要研究挑战是，标签是使用昂贵的高分辨率SPOT-6卫星数据开发的，而免费和中等分辨率的Sentinel-2数据没有等效的标签。因此，提出的工作将开发一种无监督领域自适应方法，该方法旨在从一个领域获取数据和相应的标签，并将其用于训练相关但不同领域的语义分割模型。这项工作的主要好处是推进了大规模监测影响北方驯鹿群的破碎栖息地的方法，从而支持重要的保护成果。其他好处包括推进一般无监督域自适应，从更精细的空间分辨率域到更粗糙的空间分辨率域，为其他相关环境监测任务提供更好的机会，使用免费的Sentinel-2数据，并开发基线方法以建立光谱或时间序列分析。

In Canada's vast northern region of boreal forest and wetlands, disturbances such as roads, seismic exploration, pipelines, and energy transmission corridors are a leading cause of the decline of woodland caribou (Rangifer tarandus) - boreal population. As a result, a deep understanding of these "linear disturbances" has become a research and forest management priority in Canada. An ideal tool to support managing linear disturbances is a way to automatically generate cost-effective maps that accurately identify this form of forest habitat fragmentation. As a result, the focus of this project is to develop an approach for automated map production from multispectral satellite images using deep learning methods. Specifically, convolutional neural networks designed for classifying each pixel in a satellite image will be used to produce linear disturbance maps from Sentinel-2 data. The main research challenge to be addressed is that labels were developed using expensive high-resolution SPOT-6 satellite data, and there are no equivalent labels for the free and medium-resolution Sentinel-2 data. Thus, the proposed work will develop an unsupervised domain adaptation approach, which aims to take the data and corresponding labels from one domain and adapt it for training semantic segmentation models in a related, but different domain. The main benefit of this work is advancing methods for monitoring the fragmented habitats affecting the boreal Caribou herds at scale, thereby supporting important conservation outcomes. Additional benefits include advancing general unsupervised domain adaptation from finer spatial resolution domains to coarser ones, providing improved opportunities to use freely available Sentinel-2 data for other related environmental monitoring tasks, and developing baseline methodology to build upon for spectral or time-series analysis.