Last active
April 1, 2019 07:51
-
-
Save weiji14/603a57051bbcabe8c93157dac9954e72 to your computer and use it in GitHub Desktop.
Short talk during the OBIA research group meeting at ZGIS on my Deep Learning in the Cryosphere project
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| { | |
| "cells": [ | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "slide" | |
| } | |
| }, | |
| "source": [ | |
| "# **Deep Learning in the Cryosphere**\n", | |
| "\n", | |
| "## Using deep neural networks to investigate subglacial water in Antarctica\n", | |
| "\n", | |
| "\n", | |
| "[OBIA](https://zgis.at/obia) Research Group Talk @ [ZGIS](https://zgis.at), ~11.00am-12.00noon Monday, 1 April 2019\n", | |
| "\n", | |
| "by [Wei Ji Leong](https://orcid.org/0000-0003-2354-1988), Erasmus+ Exchange PhD Student\n", | |
| "\n", | |
| "Supervisors [Dr. Huw Horgan](https://orcid.org/0000-0002-4836-0078) and Dr. Brian Anderson\n", | |
| "\n", | |
| "[Antarctic Research Centre](https://www.victoria.ac.nz/antarctic/), Victoria University of Wellington, New Zealand" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "skip" | |
| } | |
| }, | |
| "source": [ | |
| "# Outline\n", | |
| "\n", | |
| "- Research Aims\n", | |
| "- Background information\n", | |
| " - Subglacial Hydrology\n", | |
| " - Deep Neural Networks\n", | |
| "- DeepBedMap Project Initial Results\n", | |
| " - How to get a better BEDMAP for Antarctica?\n", | |
| " - Super Resolution Neural Network Model Design\n", | |
| " - Results at Pine Island Glacier focus area\n", | |
| "- Next steps and Discussion" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "# 1. Research Aims\n", | |
| "\n", | |
| "Apply **deep learning** to extract information from cryospheric **remote sensing** datasets, with a focus on Antarctic **subglacial hydrology**.\n", | |
| "The aims are as follows:" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "1. Use a **Super-image Resolution Convolutional Neural Network (SRCNN)** to increase the spatial resolution of cryospheric datasets." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "2. Locate **subglacial lakes** from **integrated geophysical/remote sensing datasets** using a Convolutional Neural Network (ConvNet)." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "3. Create a new **inventory**/**map** of Antarctic **subglacial lakes, channels and aquifers**." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "<img src=\"https://static.cambridge.org/resource/id/urn:cambridge.org:id:binary-alt:20160626060105-17421-optimisedImage-S095410201200048X_fig2g.jpg\" alt=\"Fourth Inventory of Antarctic Subglacial Lakes from Wright and Siegert 2012\" style=\"margin:0px auto;display:block\" width=\"50%\"/>\n", | |
| "\n", | |
| "379 lakes in the Fourth Inventory of Antarctic Subglacial Lakes from [Wright and Siegert 2012](https://doi.org/10.1017/S095410201200048X).\n", | |
| "Inferred from multiple data sources such as Ice-Penetrating **Radar** or **Seismic** Soundings, Surface Elevation Change using **Laser/Radar Altimetry** Satellites, etc.\n", | |
| "Much more data available since then, so it's time for an update!" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "slide" | |
| } | |
| }, | |
| "source": [ | |
| "# 2. Background\n", | |
| "\n", | |
| "- Why is the **subglacial** component important?\n", | |
| "- What are deep **neural networks**?" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "slide" | |
| } | |
| }, | |
| "source": [ | |
| "## How glaciers flow\n", | |
| "\n", | |
| "Glaciers flow via a combination of three processes:\n", | |
| "\n", | |
| "<img src=\"https://upload.wikimedia.org/wikipedia/commons/3/3a/Glacier_flow-mechanisms.png\" alt=\"The three components that contribute to glacier flow - Ice deformation, sliding at the bed, and bed deformation\" style=\"margin:0px auto;display:block\" width=\"60%\"/>" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "Of the three processes, those occuring **below the glacier** are the least understood.\n", | |
| "\n", | |
| "See Cuffey & Paterson, 2010." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## Subglacial hydrology\n", | |
| "\n", | |
| "- Water beneath a glacier plays an important role in increasing/decreasing basal slipperiness." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "- Knowing **where** water lies beneath the Antarctic ice sheet will help inform **how fast** ice will drain into the sea." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "- This feeds into the bigger picture question:\n", | |
| " - \"How quickly will **sea level rise** and affect coastal communities around the globe?\"" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## Deep Neural Networks\n", | |
| "\n", | |
| "An artificial neural network is a system made up of neurons, loosely based on biological neural networks.\n", | |
| "The 'deep' term comes when you join two or more neural layers one after another." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "<img src=\"https://screenshotscdn.firefoxusercontent.com/images/43c38c34-1ab8-44d8-b9d3-d4aa4d408855.png\" alt=\"Schematic digram of a) 1-neuron, b) a shallow 1-layer network, c) a deep 2-layer network\" style=\"margin:0px auto;display:block\" width=\"80%\"/>\n", | |
| "\n", | |
| "See also Deep Learning *Nature* review paper by [LeCun, Bengio, & Hinton, 2015](https://doi.org/10.1038/nature14539) or Deep Learning book by [Goodfellow, Bengio, & Courville, 2016](http://www.deeplearningbook.org)." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## Convolutional Neural Networks\n", | |
| "\n", | |
| "Neurons that work **spatially**.\n", | |
| "Kernels/Filters are learned to detect 'Features' like edges, corners, etc.\n", | |
| "Subsequent convolutional layers build on these basic representations to learn complex high level features like texture, patterns and shapes.\n", | |
| "\n", | |
| "<table><tr>\n", | |
| "<td><img src=\"http://deeplearning.stanford.edu/wiki/images/6/6c/Convolution_schematic.gif\" alt=\"Animation of a Convolution operation\"/></td>\n", | |
| "<td><img src=\"https://deliveryimages.acm.org/10.1145/3070000/3065386/figs/f3.jpg\" alt=\"Convolutional kernels from Krizhevsky et al. 2017\"/></td>\n", | |
| "</tr></table>\n", | |
| "\n", | |
| "Left: Animation of Convolution Operation.\n", | |
| "Right: Examples of Convolutional kernels from [Krizhevsky et al. 2017](https://doi.org/10.1145/3065386)" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "slide" | |
| } | |
| }, | |
| "source": [ | |
| "# 3. DeepBedMap Project\n", | |
| "\n", | |
| "Using a deep neural network to better resolve the bed topography of Antarctica." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## Why do we need a better BEDMAP for Antarctica?" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "- Current bed topography map (BEDMAP2) has a coarse spatial resolution of **1000m**.\n", | |
| " - Actually a 5000m grid product resampled to 1000m.\n", | |
| " - The bed of Antarctica is poorly understood compared to the surface\n", | |
| " - E.g. There is now an **8m** digital surface elevation model of Antarctica" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "- Accurate mapping of subglacial lakes requires better knowledge of the bed!" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "- Higher resolution bed will also help to capture more processes in ice sheet models." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## How to get a better BEDMAP for Antarctica?" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "- We have high resolution bed elevation maps over **some** parts of Antarctica (collected using ice-penetrating radar surveys)." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "<img src=\"https://user-images.githubusercontent.com/23487320/49407666-392e9080-f7be-11e8-9788-061dc5040796.png\" alt=\"Map of radio-echo-sounding datasets around Antarctica used in Gardner et al., 2018 paper\" style=\"margin:0px auto;display:block\" width=\"50%\"/>\n", | |
| "\n", | |
| "Figure showing Radio-echo-sounding datasets around Antarctica from [Gardner et al. 2018](https://doi.org/10.5194/tc-12-521-2018)" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "- We have high resolution surface maps over **practically all** of Antarctica (from optical, laser and radar satellites)." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "<img src=\"https://www.pgc.umn.edu/files/2018/08/REMA-hillshade-rendering-800px-768x768.jpg\" alt=\"Hillshade Map of the Reference Elevation Model of Antarctica from Howat et al. 2018\" style=\"margin:0px auto;display:block\" width=\"50%\"/>\n", | |
| "\n", | |
| "Hillshade Map of the Reference Elevation Model of Antarctica from [Howat et al. 2018](https://doi.org/10.7910/DVN/SAIK8B)" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## How about\n", | |
| "\n", | |
| "- We **train** a neural network model at the places where we have **high resolution** data.\n", | |
| " - Model learns bed topography given other high resolution surface datasets.\n", | |
| " - High resolution groundtruth areas provide 'answer' to train the neural network.\n", | |
| " \n", | |
| " `X(Surface inputs) -- function(X) --> Y(Groundtruth bed)`" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "- **Apply** the trained model to **fill in the gaps** where there is little/no survey data.\n", | |
| "\n", | |
| " `X(Surface inputs) -- function(X) --> Y(High Resolution Bed)`" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "slide" | |
| } | |
| }, | |
| "source": [ | |
| "## Neural Network Model Training Set-up\n", | |
| "\n", | |
| "<img src=\"https://yuml.me/diagram/scruffy;dir:LR;scale:180/class/[BEDMAP2 (1000m)]->[Neural Network model],[REMA (100m)]->[Neural Network model],[MEASURES Ice Flow Velocity (500m resampled)]->[Neural Network model],[Neural Network model]->[High res bed DEM (250m)],[High res bed DEM (250m)]->[Discriminator],[Groundtruth Image (250m)]->[Discriminator]\" alt=\"3 input SRGAN model\"/>" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "fragment" | |
| } | |
| }, | |
| "source": [ | |
| "Inputs inspired by this equation: $\\frac{dz}{dt} = M_b + \\nabla \\cdot u H$\n", | |
| "\n", | |
| "where change in ice surface elevation $\\frac{dz}{dt}$ is equal to mass balance $M_b$ plus divergence in velocity $\\nabla \\cdot u$ multiplied by ice thickness $H$ (surface elevation $z_s$ - bed elevation $z_b$)\n", | |
| "\n", | |
| "\n", | |
| "Currently we have MEASURES Ice Velocity $\\nabla \\cdot u$, REMA $z_s$ and BEDMAP $z_b$ used as inputs.\n", | |
| "'Assuming' steady state for elevation change $\\frac{dz}{dt}$ and mass balance $M_b$." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## Neural Network Model Architecture\n", | |
| "\n", | |
| "- Enhanced Super Resolution Generative Adversarial Network (ESRGAN) adapted from [Wang et al. 2018](https://arxiv.org/abs/1809.00219)\n", | |
| "- Currently built using Chainer (Python-based deep learning library)\n", | |
| "\n", | |
| "\n", | |
| "\n", | |
| "\n", | |
| "\n", | |
| "- Feature Extraction --> Non-linear Mapping --> High Resolution Reconstruction" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## Neural Network Training Areas\n", | |
| "\n", | |
| "Currently trained on **2499** high resolution 8x8km image tiles (with overlaps) from Pine Island Glacier, Thwaites Glacier, Rutford Ice Stream, Siple Coast, Gamburtsev Subglacial Mountains and Wilkes Subglacial Basin.\n", | |
| "\n", | |
| "<img src=\"https://user-images.githubusercontent.com/23487320/55294313-e62db880-5400-11e9-98fe-08c762ac7b83.png\" alt=\"Neural Network Training Areas around Antarctica\" style=\"margin:0px auto;display:block\" width=\"60%\"/>" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "slide" | |
| } | |
| }, | |
| "source": [ | |
| "# Focus Area - Pine Island Glacier\n", | |
| "\n", | |
| "- Fantastic high resolution grids from [Ice Sheet Stability Programme (iSTAR) surveys](https://istar-gis-features.data.bas.ac.uk/)" | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "<table><tr>\n", | |
| "<td><img src=\"https://media.springernature.com/lw900/springer-static/image/art%3A10.1038%2Fs41467-017-01597-y/MediaObjects/41467_2017_1597_Fig1_HTML.jpg\" alt=\"3D view of high resolution radar surveys at Pine Island Glacier from Bingham et al. 2018\" style=\"margin:0px auto;display:block\" width=\"80%\"/></td>\n", | |
| "<td><img src=\"https://media.springernature.com/lw900/springer-static/image/art%3A10.1038%2Fs41467-017-01597-y/MediaObjects/41467_2017_1597_Fig2_HTML.jpg\" alt=\"Comparing BEDMAP2 to high resolution radar surveys at Pine Island Glacier from Bingham et al. 2018\" style=\"margin:0px auto;display:block\" width=\"100%\"/></td>\n", | |
| "</tr></table>\n", | |
| "\n", | |
| "High resolution radar surveys of Pine Island Glacier from [Bingham et al. 2017](https://doi.org/10.1038/s41467-017-01597-y)." | |
| ] | |
| }, | |
| { | |
| "cell_type": "markdown", | |
| "metadata": { | |
| "slideshow": { | |
| "slide_type": "subslide" | |
| } | |
| }, | |
| "source": [ | |
| "## Inputs into trained Neural Network Model\n", | |
| "\n", | |
| "
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
https://nbviewer.jupyter.org/format/slides/gist/weiji14/603a57051bbcabe8c93157dac9954e72#/?flush_cache=true