[CS] [masters-presentation] Kurihana/MS Presentation/Dec 15, 2020

[Tricia Baclawski]

This is an announcement of Takuya Kurihana's MS Presentation.


https://uchicago.zoom.us/j/97465451666?pwd=V0xEQmk3Mm5mNVBHVFFlSFB3UmdOdz09
Password: 767407

------------------------------------------------------------------------------
Date:  Tuesday, December 15, 2020

Time:  3:00 PM

Place:  via zoom

M.S. Candidate:  Takuya Kurihana

M.S. Paper Title: A Data-Driven Novel Cloud Classification Framework
Based on An Efficient Cloud Representation via Rotational And
Non-Rotational Autoencoder

Abstract:
Clouds are one of the largest sources of uncertainty in projections
from global climate models. One promising solution to this problem is
to leverage the large observational datasets produced by satellite
instruments over several decades to improve understanding of cloud
dynamics and feedbacks. However, due to the large size and diversity
of these datasets, conventional statistical methods are in general
poorly suited to extracting the features that are commonly used to
distinguish among different standard cloud classes, which furthermore
are often overly restrictive in their definitions, particularly when
it comes to classifying continuous, intermediate, and complex clouds.
In this thesis we explore a new approach to cloud classification in
which an unsupervised deep learning framework is used to automate the
classification of cloud patterns and textures without any assumptions
concerning artificial cloud categories. Specifically, we leverage deep
neural network autoencoder methods to map images produced by satellite
instruments into a compact latent representation, to which we then
apply hierarchical agglomerative clustering to identify novel cloud
types. Our convolutional autoencoder uses a rotation invariant loss
function to achieve rotation invariance, meaning that it maps images
that differ only in the rotational orientation to the same latent
representation; this important feature means that comparable cloud
images appearing in different orientations (e.g., due to different
orographies) are correctly mapped to the same class. We apply our new
method to radiance data from NASA's Moderate Resolution Imaging
Spectroradiometer (MODIS) instrument on the Terra
satellite-specifically, to MODIS level 1B calibrated radiances (MOD02
bands 6, 7, 20, 28, 29, and 31) and level 2 cloud masks (MOD35) and
apply a suite of evaluation protocols to assess its effectiveness. To
validate that our autoencoder is indeed rotationally invariant we
apply smoothing and scrambling protocols to verify that it learns
rotationally invariant and spatial features, and compare its
performance against variants that are not rotation invariant. To
validate that our autoencoder yields physically meaningful cloud
classes, we evaluate whether the resultant cloud classes are well
correlated with physical metrics. We find that our classes produce
appropriately spatially coherent classifications and capture
meaningful aspects of cloud physics without being reproducible from
mean values of physics parameters alone. Our results support the
possibility of using unsupervised data-driven frameworks for automated
cloud classification and pattern discovery without requiring the prior
hypothesis of ground-truth labeled data. It lays the groundwork for
expanded studies in which these methods will be applied to all 19
years (2000-2018) of MODIS data, and also to other datasets.

Takuya's advisor is Prof. Ian Foster

Login to the Computer Science Department website for details:
 https://newtraell.cs.uchicago.edu/phd/ms_announcements#tkurihana

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Tricia Baclawski
Student Affairs Administrator
Computer Science Department
5730 S. Ellis - Room 350
Chicago, IL 60637
pbaclawski at uchicago.edu
(773) 702-6854
/pronouns: she, her, hers/
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=