A machine learning approach to constraining ice volume and potential loss in High Mountain Asia

项目背景。亚洲高山地区(HMA)的冰川正在经历大规模消融[1]，这对依赖冰川获取关键水资源的数亿人产生了影响[2]。对喜马拉雅冰川物质平衡和相关径流的可能轨迹的预测非常不确定--部分原因是缺乏冰川厚度的知识，这决定了冰川对气候变化的反应[3]。随着该区域90,000个冰川的遥感记录不断增加[例如，4]，有可能计算区域厚度并模拟冰川对气候变化的反应[5]，但到目前为止，几乎没有可用的测量来限制厚度模型。随着喜马拉雅山脉覆盖昆布盆地冰川的首次航空冰厚测量的完成，这些模型终于可以得到约束了。该项目将通过将新的现场和卫星数据产品与先进的建模和机器学习方法相结合，研究HMA冰川对气候变暖的敏感性。更具体地说，1.经过ML训练的模型能否同化/反演来自卫星数据的HMA厚度数据？2.现场观测如何提供信息并改进这种反模型？3.改进的冰川厚度和消融评估如何有助于模拟亚洲冰川未来响应气候变化的行为？方法论。推断厚度的方法将基于[5]的蟒蛇同化框架，该框架利用指示冰川模型，这是一个深度学习仿真器[7]。该框架已成功地应用于阿尔卑斯山冰川，但不适用于观测类型和可获得性不同的HMA冰川。PHD的工作将包括修改应用于HMA冰川的框架；根据潜在的2级和3级EO数据集准备和试验输入：海拔变化(基于WorldView[8]和ASTER[9]和Cryosat[1]的数据)；以及海拔和冰川速度(ITS_LIVE)。选定冰川的空中厚度测量将由BAS监督员提供，以便对方法进行验证和改进。重要的是，由于政府间冰川机制仅在阿尔卑斯山冰川上进行了培训和应用，其性能还将在一小部分冰川物理模型[10]上进行测试，有可能通过进一步的深入学习来改进政府间全球机制。厚度增加对未来冰川损失的影响将通过使用IGM的十年建模来检验。上下文：这个博士项目将与大融化项目密切合作，大融化项目是最近由BAS领导的、跨机构的NERC重点专题赠款，旨在填补全球山区水资源知识的关键空白，但不包括新的ML厚度估计方法。该博士的努力将提供给大融化项目并向其提供信息，学生将强烈参与项目会议和讨论，使之能够与BAS、利兹和CEH的科学家进行强有力的互动，超越监督团队和行业合作伙伴。

Project Background. Glaciers in High Mountain Asia (HMA) are experiencing mass loss [1], with implications for the hundreds of millions of people who depend on them for critical water resources [2]. Projections of the likely trajectory of Himalayan glacier mass balance, and associated runoff, are highly uncertain - due in part to lack of knowledge of glacier thickness, which determines glacier response to climate change [3]. With an ever-growing remote-sensing record for the 90,000 glaciers in the region [e.g., 4], there is potential to compute thicknesses regionally and model glacier response to climate change [5], but until now, very few measurements were available to constrain the thickness models. With the completion of the first airborne [6] ice-thickness survey in the Himalayas, covering the glaciers of the Khumbu basin, these models can finally be constrained. This project will investigate HMA glacier sensitivity to climate warming by combining new field and satellite data products with advanced modelling and machine learning methods. More specifically, 1. Can ML-trained models assimilate/invert for HMA thickness data from satellite data?2. How do field observations inform and improve such inverse models?3. How does the improved assessment of glacier thickness and ablation aid in modelling the future behaviour of Asian glaciers in response to climate change?Methodology. The method for inferring thickness will be based around the python assimilation framework of [5], which makes use of the Instructed Glacier Model, a deep learning emulator [7]. The framework has been applied successfully to Alpine glaciers, but not to HMA glaciers where type and availability of observations differs. The work of the PhD will involve modifying the framework for application to HMA glaciers; preparing and experimenting with inputs based on potential Level-2 and Level-3 EO datasets: elevation change (WorldView [8] and ASTER [9] and Cryosat [1] based data);as well as elevation and glacier velocities (ITS_LIVE). Airborne thickness measurements of select glaciers will be provided by BAS supervisors, allowing validation and refinement of the methodology. Importantly, as IGM has only been trained on and applied to Alpine glaciers, its performance will also be tested on a small subset against a physical glacier model [10], with potential to improve the IGM through further deep learning. The impacts of the improved thickness on future glacier loss will be examined through multidecadal modelling using the IGM.Context: This PhD project will engage strongly with The Big Thaw, a recently-funded BAS-led, cross-institutional NERC Highlight Topics grant which aims to fill key gaps in knowledge of global mountain water resources, but does not encompass novel ML approaches to thickness estimation.The efforts of this PhD will feed into and inform The Big Thaw, and the student will be strongly involved in project meetings and discussions, enabling strong interaction with scientists at BAS, Leeds, and CEH that extend beyond the supervisory team and industry partner.