Ji yu shen du xue xi yi ji duo shi xiang yao gan ying xiang de Gelinglan dao bing chuan qian yuan bian hua yan jiu

Quantification of front variations of glaciers in Greenland is essential for under-standing ice-ocean interaction, glacier dynamics, and a more accurate projection of global sea-level rise. Most previous studies manually delineated ice fronts from remote sensing images. Such manual efforts, however, are becoming more labor- intensive and even impractical for quantifying temporal changes of ice fronts at a large number of glaciers in a detailed manner, because the volume and diversity of satellite images have been increasing significantly.Automating the identification of glaciological features such as calving fronts therefore becomes necessary. In the first study, we applied U-Net, a deep learning architecture, to multi-temporal Synthetic Aperture Radar images taken by the TerraSAR-X satellite and automated the delineation of the calving front positions of Jakobshavn Isbræ from 2009 to 2015. The results are consistent with the manually delineated products generated by the Greenland Ice Sheet Climate Change Initiative project with the mean test errors of 38 meters ( 6 pixels). We show that the calving fronts of Jakobshavn’s two main branches retreated at mean rates of −117 ± 1 m yr⁻¹ and −157 ± 1 m yr⁻¹, respectively, during the years 2009 to 2015. We suggest that the retreat of the calving front into an overdeepened basin whose bed is retrograde may have accelerated the retreat after 2011, while the inland-uphill bed slope behind the bottom of the overdeepened basin has prevented the glacier from retreating further after 2012. Although we demonstrate the feasibility of deep-learning-based method, this study is limited to a specific study area (e.g., Jakobshavn Isbræ) and dataset (e.g., TerraSAR-X).The second study further improves the generalization and robustness of the deep- learning-based method to release its constraint of applicable region and data. The method can be applied an outlet glacier or remote sensing datasets that are not included in training, and we integrate seven remote sensing datasets into a single deep learning network. The core datasets include optical (Landsat-8 and Sentinel-2) and synthetic aperture radar images (Envisat, ALOS-1, TerraSAR- X, Sentinel-1, and ALOS-2) taken over Jakobshavn Isbræ, Kangerlussuaq, and Helheim, spanning from 2002 to 2019. We evaluate four neural network architectures (e.g., U-Net, DeepLabv3+ with Resnet, DRN, and Mobilenet as the backbones) and three pre-processing procedures (e.g. histogram normalization, linear stretching, and no histogram manipulation). We find that the combination of histogram normalization and DRN-DeepLabv3+ has the lowest mean test error, 86 meters. We produce a sub-weekly calving front dataset for the three glaciers by integrating seven datasets. Owing to the high temporal resolution, our results reveal large calving events as well as synchronous and asynchronous front variations between neighboring branches or glaciers.In summary, by applying deep-learning-based method to multi-temporal and multi-sensor remote sensing imagery, I successfully automate the delineation of front positions of Jakobshavn Isbræ, Kangerdlugssuaq, and Helheim. Due to the transferable nature of deep learning, the methodology can be applied to many other tidewater glaciers both in Greenland and elsewhere in the world. Our method has the potential to produce a spatially, temporally comprehensive calving front dataset, which will help to improve the understanding of front dynamics and glacial mass balance.冰川前緣變化的定量研究對於理解海冰交互作用，冰川動力學，以及更精準地預測全球海平面上升有著重要作用。先前的研究大多數基於從遙感影像中人工提取邊界。然而，此種方法耗時耗力，而且由於衛星影像的數量以及多樣性的急劇增加，人工提取無法實現對於大量冰川的細緻定量分析。自動化諸如冰川前緣的冰川學特徵提取就變的極為重要。在第一個研究中，我們將名為U-Net的深度學習神經網絡應用於多時項TerraSAR-X合成孔徑雷達影像上，並從中自動提取了Jakobshavn Isbræ冰川自2009至2015年間的前緣邊界。我們的結果和Greenland Ice Sheet Climate Change Initiative項目人工提取的產品基本一致，並且結果的測試誤差在38米（約6個像素點）。結果顯示從2009到2015，Jakobshavn Isbræ冰川的兩個分支的平均回退速度分別為-117±1米每年以及-157±1米每年。我們認為在2011之後年冰川會退到一個冰床向內陸凹陷的盆地，並導致了其回退的加速。在2012年之後冰川邊界處的冰床向內陸上升，阻止了冰川繼續會退。儘管我們的研究證實了深度學習方法的可行性，該方法只適用於特定區域（如Jakobshavn Isbræ）以及特定數據集（如TerraSAR-X）。第二項研究進一步增加了深度學習方法的普適性以及穩健性並解除了其對於特定區域和數據集的限制。該方法可以應用於訓練數據所不含蓋的冰川以及數據集。我們的核心數據包括從2002到2019，覆蓋Jakobshavn Isbræ，Kangerlussuaq和Helheim的光學（Landsat-8和Sentinel-2）以及合成孔徑雷達（Envisat，ALOS-1，TerraSAR-X，Sentinel-1和ALOS-2 ）影像。我們評估了四個深度學習網絡（如U-Net, 以Resnet，DRN，Mobilenet為核心的DeepLabv3+）以及三個預處理方法（直方圖正態化，線性拉伸和無處理）。我們發現DRN-DeepLabV3+和直方圖正態化的組合有最好的準確度，其測試誤差為86米。通過整合七個數據集，冰川前緣數據的採樣頻率超過了每週兩次。得益於數據的高時間分辨率，我們的時間序列可以直接顯示大的冰裂事件，並且可以直觀顯示出臨近冰川或分支前緣的同異步變化。總的來說，我們通過將深度學習算法應用於多時項，多感應器遙感影像，成功地自動提取了Jakobshavn Isbræ，Kangerdlugssuaq，和Helheim冰川的前緣。由於深度學習自身的可移植性，我們的方法還能應用於格陵蘭島乃至世界其他地區的冰川。我們的方法有潛力提供一個在時間、空間尺度的全面性達到前所未有高度的冰川前緣數據集。該數據將會對理解冰川前緣動態學以及冰川質量平衡提供有力幫助。Zhang, Enze.Ph.D. Chinese University of Hong Kong 2020.Includes bibliographical references (leaves 86-99).Abstracts also in Chinese.Title from PDF title page (viewed on September 13, 2022).Zhang, Enze