Deep multi-scale learning for automatic tracking of internal layers of ice in radar data

Abstract

In this study, our goal is to track internal ice layers on the Snow Radar data collected by NASA Operation IceBridge. We examine the application of deep learning methods on radar data gathered from polar regions. Artificial intelligence techniques have displayed impressive success in many practical fields. Deep neural networks owe their success to the availability of massive labeled data. However, in many real-world problems, even when a large dataset is available, deep learning methods have shown less success, due to causes such as lack of a large labeled dataset, presence of noise in the data or missing data. In our radar data, the presence of noise is one of the main obstacles in utilizing popular deep learning methods such as transfer learning. Our experiments show that if the neural network is trained to detect contours of objects in electro-optical imagery, it can only track a low percentage of contours in radar data. Fine-tuning and further training do not provide any better results. However, we show that selecting the right model and training it on the radar imagery from the start yields far better results.

Subglacial lakes and their changing role in a warming climate

“…In particular, the sudden drainage of lakes beneath mountain glaciers poses a hazard to residents and infrastructure in valleys below 124,246 and so a better understanding of water storage and drainage beneath glaciers in populous areas and how the risk might change due to climate warming should be a priority. Improvements in the spatial and temporal coverage and resolution of satellites (e.g., Sentinel-1, CryoSat-2, Sentinel-3 and ICESat-2) 7,89,92,93 coupled with developments in lake detection automation 82 and machine learning 247 will likely allow these gaps to be filled, particularly for lakes that are smaller and traditionally more difficult to detect. The transition from a set of individual satellites to systems dominated by fleets and constellations of many satellites, will enable increased availability of high-resolution multi-temporal DSMs 97,248 , which are particularly well suited for picking out smaller lakes 55,59,249 .…”

Long Time-Series Glacier Outlines in the Three-Rivers Headwater Region From 1986 to 2021 Based on Deep Learning

“…Among them, the eight methods performed best in testing area 1, almost the same as in scene 1, which indicates that the eight methods have good generalization abilities when the testing area is highly similar to the training area. In fact, most previous studies verified or compared the performances of deep learning models using the dataset division method used in scene 1 and in testing area 1 in scene 2 [26], [60], [61]. In testing area 2, when the glaciers were very clear but the surrounding geographic environment of the glaciers was quite different from that in the training area, the performances of the eight methods decreased to a certain extent.…”

Deep Learning on Airborne Radar Echograms for Tracing Snow Accumulation Layers of the Greenland Ice Sheet

“…We collectively refer to the above degradation phenomena as "noise". This noise makes tracking accumulation layers challenging even for experienced glaciologists and significantly increases the workload for conventional vision algorithms [3], [5]. Even convolutional neural networks (CNNs) [6], which have become the standard machine learning algorithms for image processing, have issues in making accurate predictions from these images [5], [7], [8] since any interruptions to their input can drastically affect their prediction capability [9].…”

“…This noise makes tracking accumulation layers challenging even for experienced glaciologists and significantly increases the workload for conventional vision algorithms [3], [5]. Even convolutional neural networks (CNNs) [6], which have become the standard machine learning algorithms for image processing, have issues in making accurate predictions from these images [5], [7], [8] since any interruptions to their input can drastically affect their prediction capability [9]. Further, as the radar images are noisy, visually picking the layers for manual annotations either creates missing or incomplete labels for most firn layers in these images [3].…”

“…In this scenario, a possible solution is to have a neural network which can reduce the inherent noise present in images while simultaneously extract-ing features from them. Such a hybrid architecture would be optimum for highlighting firn layers in radar echograms, which can be used for calculating snow accumulation rates [3], [5].…”