Remote Sensing Scene Classification

遥感是一门从空中或卫星获取地球数据的科学，为智能农业实践、北方沃茨航运和温室气体监测等重要工作提供了关键信息。产生了大量的数据，迫切需要基于计算机的算法来解释数据。例如，利用加拿大雷达卫星-II，加拿大冰服务人员每年人工判读4000个场景，提供大面积（500公里乘500公里）区域的冰类型图，推断冰的厚度和强度，以促进航运和破冰的决策支持系统。为了节省时间和金钱并减少人为偏见，需要能够分析这些图像的机器算法。我特别感兴趣的是这种基于雷达的海冰图像的自动化解释，被认为是遥感问题的“圣杯”，由于传感器特性和环境变化混淆了冰的类型，这是一个挑战。 我的研究团队将通过使用操作员的粗略地图来初始化自动化解决方案以识别任何冰类型，从而在这一目标上取得重大进展。例如，虽然人工粗略的解释将表明一个大区域包含30%的厚冰，但我们的传感器分辨率分类将精确定位厚冰的位置。我们的原型已通过MDA（RADARSAT-II业主）和CIS的验证。 现在，在与MDA的合作中，我们正在积极研究一种可操作的CIS实施方案，以在没有操作员输入的情况下区分冰和开放水域。我们生成的详细地图预计将成为船舶在危险沃茨航行的必要条件（以降低风险并节省时间和燃料），并为气候模型提供输入。其他北方国家和石油公司将有兴趣购买从所提出的算法得出的地图。我在计算机视觉方面的研究已经并将继续用于其他应用，如医学成像和视频解释。

Remote sensing, the science of aerial or satellite data capture of the earth, provides crucial information for such important endeavours as smart agricultural practices, shipping in northern waters, and monitoring of greenhouse gases. Vast amounts of data are generated, and computer-based algorithms to interpret the data are urgently needed. For example, using the Canadian RADARSAT-II satellite, Canadian Ice Service (CIS) personnel manually interpret 4000 scenes annually, providing ice type maps over huge (500km by 500km) regions that infer ice thickness and strength to facilitate decision support systems for shipping and icebreaking. To save time and money and to reduce human bias, there is a demand for machine algorithms that can analyze this imagery. My particular interest is the automated interpretation of this radar-based sea ice imagery, considered the "holy grail" of remote sensing problems, a challenge due to the sensor characteristics and environment variability that confuse the ice types. My research team will make significant advances toward this objective by using an operator's coarse map to initialize an automated solution to identify any ice type. For example, whereas manual coarse interpretation would indicate a large region as containing 30% thick ice, our sensor-resolution classification would pinpoint the thick ice location. Our prototype has been validated by MDA (RADARSAT-II owner) and CIS. Now, in co-operation with MDA, we are actively investigating an operational CIS implementation to distinguish ice from open water without operator input. The detailed maps we generate are expected to be a necessity for ship navigation in hazardous waters (to reduce risk and save time and fuel) and to provide input into climate models. Other northern nations and oil companies would be interested in purchasing maps derived from the proposed algorithm. My research in computer vision required to solve this problem has been and will continue to be used in other applications, such as medical imaging and video interpretation.