Slides | Steefan Contractor

Aceas Multiage Seaice Classification

Thu, 11 Jul 2024 11:30:37 +1000

Bayesian updates to multi-age Antarctic sea-ice concentrations using GNSS-R data

Speaker: Steefan Contractor

Coauthors: Shane Keating (UNSW), Jessica Cartwright (Spire Global), Alex Fraser (UTAS)

Sea Ice

Floating ice that forms from the freezing of seawater
Melting and formation of sea ice affects the ocean salinity and heat content
The changes in ocean density and temperature affect the ocean circulation as evidenced by recent coverage on AMOC weakening
It affects the Earth’s energy balance by reflecting ten times more sunlight compared to water
It acts like a blanket affecting not just heat exchange between the ocean and atmosphere but also gases
Through the changes in polar air masses it also affects atmospheric circulation

Image sources:

Young Ice (YI)

Newly formed ice
can be rough or smooth
less than 30cm thick

First-year Ice (FYI)

Ice that has survived one summer melt season
can be level, rough or have ridges
30cm to 2m thick

Multi-year Ice (MYI)

Ice that has survived more than one summer melt season
typically smoother than FYI
Over 2.5m thick and hence protrudes above the waterline
has extremely low salinity compared to YI and FYI

Remote Sensing of Sea Ice

Active sensors

Radar altimeters (CrysoSat-2)
Laser altimeters (IceSat-2)
Scatterometers (Metop A/B/C - ASCAT)
Synthetic Aperture Radar (Radarsat-2)

Passive sensors

Passive microwave radiometers (SMOS)

Hybrid sensors

Global Navigation Satellite System Reflectometry - GNSS-R

IUP U. Bremen Multiage Ice Concentration

GNSS-R

Ok here’s how it works: there is a direct and reflected signal
we aggregate the intensity of each signal into time delay (phase space) and doppler shift (frequency space) bins giving us these DDMs
from these DDMs we get the various variables that we will use as features to classify the ice types
- power: integrating over the DDM space we get the power
- reflectivity: ratio of the reflected to the incident power
- snr: subtract out the noise floor where we don’t see a signal and divide it by the noise floor
- phase noise: standard deviation of the phase of the signal
- excess phase noise: std dev of the phase after removing the coherent SNR component of the signal, leaving only the component of phase noise due to the surface roughness
These variables were hand picked by Spire’s domain experts (Jessica Cartwright) after lots of internal testing and this was the data that was given to us
Note that the reason we focus on these variables instead of using the entire DDMs is because the dataset is already over 10Gb with only 14 numbers per observation, let alone a whole image per observation.
There are a lot of samples, the raw data was in TB.

IUP Multiyear ice concentration and other sea ice types, Version AQ2 (Antarctic)

Provides YI, FYI and MYI concentrations
Developed by Institute of Environmental Physics, University of Bremen
Uses passive microwave (AMSR2) and scatterometer (ASCAT instruments on Metop A/B/C) data to derive initial estimates
Corrects the initial estimates using 2m surface air temperature and sea ice drift data
12.5km x 12.5km grid resolution

Melsheimer, Christian; Spreen, Gunnar; Ye, Yufang; Shokr, Mohammed (2019): Multiyear Ice Concentration, Antarctic, 12.5 km grid, cold seasons 2013-2018 (from satellite). PANGAEA, https://doi.org/10.1594/PANGAEA.909054

Sea Ice Signal in GNSS-R features?

Some details

PCA data:
- water - total ice concentration = 0%
- ice - total ice concentration > 99%
PCA features:
- reflectivity1
- snr_reflected1
- power_reflected1
- phase_noise1
- excess_phase_noise1
total explained variance: 99.25%

Some details

PCA data:
- YI - YI ice concentration > 90%
- FYI - FYI ice concentration > 99.9%
- MYI - MYI ice concentration > 99%
PCA features:
- reflectivity1
- snr_reflected1
- power_reflected1
- phase_noise1
- excess_phase_noise1
total explained variance: 99.58%

Correlation amongst GNSS-R features and ice types

Geographic distribution

= 80%25.png" height="320px">

National Snow and Ice Data Center, Boulder, Colorado USA. https://nsidc.org/data/seaice_index, last access: 2024-07-16

J. C. Comiso, A. C. Bliss, R. Gersten, C. L. Parkinson, and T. Markus (2024), Current State of Sea Ice Cover, https://earth.gsfc.nasa.gov/cryo/data/current-state-sea-ice-cover, last access: 2024-07-16.

Ice concentrations as probabilities

the spatial footprint of GNSS-R grid is around 2.5% of the size of the IUP grid
interprit ice concentration as the probabiliy of seeing that kind of ice inside that grid
assume that the ice concentrations can be erroneous
assume higher ice concentrations are more likely to be correct
semi-supervised approach: learn a mapping from GNSS-R to ice labels where IUP concentrations are close to 100%
update the IUP concentrations using the learned mapping

Bayesian update

$$ \begin{align*} P(true\, label\, |\, model\, pred) &= \frac{P(model\, pred\, |\, true\, label) \times P(true\, label)}{P(model\, pred)} \\ &= \frac{P(model\, pred\, |\, true\, label) \times P(true\, label)}{\sum_{true\, labels}{P(model\, pred\, ,\, true\, label)}} \\ &= \frac{P(model\, pred\, |\, true\, label) \times P(true\, label)}{\sum_{true\, labels}{P(model\, pred\, |\, true\, label) \times P(true\, label)}} \end{align*} $$

Two ways to get $P(model\ pred\ |\ true\ label)$:

Using a tabular ML model - model calibration required
- calibrate using test dataset and Bayes rule again

$$ P(pred\ |\ true) = \frac{P(true\ |\ pred)P(pred)}{\sum_{pred\ labels}{P(true\ |\ pred)P(pred)}} $$

Using Robust Mixture Discriminent Analysis (RMDA)

Bouveyron, C., & Girard, S. (2009). Robust supervised classification with mixture models: Learning from data with uncertain labels. Pattern Recognition, 42(11), 2649–2658. https://doi.org/10.1016/j.patcog.2009.03.027

Which ML model?

Decision tree ensemble methods

Random Forest: Fits an ensemble of N trees on random subsets (known as bagging) of data and predicts with a majority vote.

Adaptive Gradient Boosting: Trains N trees sequentially, each tree correcting the errors of the previous tree by weighting the data points that were misclassified. This is known as boosting.

Gradient Boosting: Also a boosting method in that it trains N trees sequentially, however, instead of weighting the data points, it fits each tree to the residuals of the previous tree.

LightGBM

Key improvements over vanilla gradient boosting:

Histogram-based splitting: LightGBM bins the data points into discrete bins and then splits the bins instead of the data points. This reduces the complexity of the model and speeds up training.

Leaf-wise growth: Instead of growing the tree level-wise, LightGBM grows the tree leaf-wise. This reduces the number of nodes in the tree and hence the complexity of the model.

Gradient-based One-Side Sampling: LightGBM samples the data points based on the gradient of the loss function. This speeds up training by focusing on the data points that are more informative.

Exclusive Feature Bundling: LightGBM bundles exclusive features together to reduce the number of features that need to be considered during training.

Gaussian Mixture Discriminant Analysis

\[ \begin{align*} G &\sim Discrete(K) \\ C &\sim Discrete(L) \\ X &\in \mathbb{R}^p \\ p(x|C=i) &= \sum_{j=1}^{K} P(C=i,G=j)p(x|G=j) \\ p(x) &= \sum_{i=1}^L\sum_{j=1}^{K} P(C=i,G=j)p(x|G=j) \\ &= \sum_{i=1}^L\sum_{j=1}^{K} \pi_{ij}\phi(x;\mu_ij,\Sigma_ij) \end{align*} \] Here $\pi_{ij}$ is the mixing coefficient such that $\sum_{j=1}^K \pi_{ij} = 1$, and $\phi(x;\mu_ij,\Sigma_ij)$ is the multivariate Gaussian distribution with mean $\mu_ij$ and covariance $\Sigma_ij$.

Robust Mixture Discriminant Analysis

\[ \begin{align*} p(x) &= \sum_{i=1}^L\sum_{j=1}^{K} P(C=i,G=j)p(x|G=j) \\ &= \sum_{i=1}^L\sum_{j=1}^{K} P(C=i|G=j)P(G=j)p(x|G=j) \\ &= \sum_{i=1}^L\sum_{j=1}^{K} r_{ij}\pi_j\phi(x;\mu_j,\Sigma_j) \end{align*} \] Since $\pi_j$ does not depend on $C$, we can fit a gaussian mixture model (unsupervised) to get $\pi_j$. To get the, $L\times K$, matrix $R=(r_{ij})$ parameters, we maximise the log likelihood: \[ l(R) = \sum_{i=1}^{L}\sum_{x\in \mathcal{C}_i} log(R_i\Psi(x)) \] where $\Psi(x) = (P(S=1|X=x),P(S=1|X=x),...,P(S=K|X=x))^t$ and $\mathcal{C}_i=\{x_l\}$ such that $x_l$ belongs to class $C=i$, w.r.t. $r_{ij}\in [0,1], \forall i\in {1,...,L}\text{ and }j\in{1,...,K}$, subject to $\sum_{i=1}^{L} r_{ij}=1, \forall j=1,...,K$.

Class rebalancing

Training score: 83.7%
Validation score: 81.3%

	YI	FYI	MYI	water
YI	0.59237875	0.27944573	0.0369515	0.09122402
FYI	0.02884615	0.94346154	0.01884615	0.00884615
MYI	0.08886389	0.64904387	0.22497188	0.03712036
water	0.02849003	0.01745014	0.00747863	0.9465812

The entire dataset contains 7.39M rows.
After filtering rows where we have high confidence in labels (YI>90%, FYI>99.9%, MYI>99.%,Water=100%), we are left with:
- YI: 9801 rows
- FYI: 28266 rows
- MYI: 9805 rows
- Water: 3.18M rows

Solution: SMOTE based class rebalancing

SMOTE

Synthetic Minority Over-sampling Technique
1. Identify the minority class
2. Find its nearest neighbours
3. Generate synthetic samples by interpolating between the minority class and its neighbours
After SMOTE each class contains 3.18M rows
So we randomly undersampled and treated the number of samples in each class as a hyperparameter
More sophisticated undersampling techniques (Tomek, Edited Nearest Neighbours) did not prune the dataset enough

Performance after class rebalancing


Number of resampled rows in each class	~1.6M
Training accuracy	99.34%
Validation accuracy	99.12%
Test accuracy	96.34%

Test Confusion matrix

	YI	FYI	MYI	water
YI	0.67241379	0.15922921	0.04158215	0.12677485
FYI	0.04790419	0.84677703	0.08559352	0.01972526
MYI	0.046875	0.25878906	0.64257812	0.05175781
water	0.01741254	0.00994283	0.00626461	0.96638002

Updated ice probability density

UMAP (Uniform Manifold Approximation and Projection)

UMAP is a non-linear dimensionality reduction technique that is particularly well-suited for visualizing complex data in a low-dimensional space
UMAP assumes that high-dimensional data lies on a low-dimensional manifold embedded in the higher-dimensional space

UMAP first constructs a weighted k-nearest neighbor graph from the high-dimensional data
UMAP then defines a low-dimensional representation and uses stochastic gradient descent to optimize the layout of the low-dimensional points, preserving the structure of the high-dimensional graph as closely as possible

It thus preserves both local and global structure of the data
Supervised mode:
- During knn graph construction, UMAP uses the labels to weight the edges in the graph so that points with the same label are closer in the low-dimensional space
- Also adds a loss term to the optimization function that penalizes points with the same label being far apart in the low-dimensional space

Performance after UMAP feature transformation


Number of resampled rows in each class	~300K
Training accuracy	99.99%
Validation accuracy	99.84%
Test accuracy	93.34%

Test Confusion matrix

	YI	FYI	MYI	water
YI	0.95866935	0.02116935	0.0141129	0.00604839
FYI	0.01243201	0.95648796	0.02175602	0.00932401
MYI	0.01208459	0.02819738	0.95166163	0.00805639
water	0.04524251	0.00951845	0.02106578	0.92417327

Updated ice probability density

Updated geographic distribution - winter

Updated geographic distribution - summer

GNSS-R is an exciting new data source for monitoring sea ice at very high resolution
we fitted models to predict ice types and used these models to update existing ice concentration datasets
we pick up many more locations previously missed thus improving existing ice concentration datasets
this was only one year of data, the accuracy will improve with more data

More validation required! We are open to ideas

Thank you

Deep Sea Talk

Tue, 28 Nov 2023 11:11:50 +1100

Observing oceans from above

Dr. Steefan Contractor

A lot happens on the ocean surface

Interaction with the atmosphere
- air sea heat fluxes
Interaction with the biosphere
- confluence of birds, fish and mammals
Fisheries, transport, tourism and other commercial activities
Political and Defence related activities

Sea Ice

Detecting sea ice from space

UK-Aus Space bridge Grant

GNSS-R Features

Phase Noise

Excess Phase Noise

water/ice and ice type classification

Thank you

Intro

Fri, 17 Nov 2023 17:22:03 +1100

Year 10 Data Science Work Experience Week

Welcome!

Acknowledgement of country

UNSW Kensington Campus is located on the unceded lands of the Bedigal people. We pay our respects to their Elders, past and present, as the Traditional Custodians of this land.

Health and Safety Induction

We are following the HS414 Visitors to UNSW Facilities Guideline
Visitors will be met by the person organizing the visit
Visitors need to be accompanied and supervised by UNSW staff members or teaching assistants
Visitors are not allowed in medium-or high-risk areas
Please fill the HS630 Visitor Induction Form with your contact details and check all items that apply to the induction
All visitors must agree to follow any reasonable instruction in relation to health and safety.
Report all work related hazards, incidents, injuries and illnesses to supervising staff members or teaching assistants

General safety in the classroom

No special gear is required
No food or drinks in the classroom
Please keep the classroom clean and tidy
Locate first aid equipment
Locate fire extinguishers

You are responsible for the safety of you and others around you - Take care!

if you see something unsafe, tell your teacher or other staff member

Cold & Flu safety

You are welcome to use face masks and are even encouraged to use them
Sanitise/wash hands regularly
if you feel unwell, please stay at home

Emergency Procedure

In case of an emergency

call UNSW Security Services
do not call 000

Security Services

in an emergency 9385 6666
everything else 9385 6000
Security Office located at Gate 2, open 24/7

Emergency Evacuation

BEEP BEEP: prepare to evacuate. Do not leave until WHOOP WHOOP alarm
WHOOP WHOOP: evacuate immediately. Follow staff and building warden instructions
Assemble infront of John Clancy auditorium

First Aid - Physical

Location of first aid kit and first aid room (Rm G003, E26)
Location of Automatic External Defibrillator
Any incident requiring the use of first aid, however minor, must be reported online to UNSW

First Aid - Mental

UNSW First Responders are students and staff who are trained to offer you confidential support. They understand that reporting gendered violence can be difficult and can provide you with guidance and support.

Scott Sisson, uDASH director

You can also contact a certified ally@UNSW in the School of Maths and Stats, Faculty of Science. The ally@UNSWnetwork aims to ensure UNSW is a safe and welcoming place for all LGBTIQ+ students and staff.

Map by MazeMap

UNSW Year 10 Data Science Work Experience Week

Aim

Learn the basic principles of data science. Exploratory data analysis, statistical modelling and visualisation.

Students will

learn basic programming concepts
get hands-on experience in analysing real-world datasets
work in independent teams on data science projects

UNSW Data Science Hub (uDASH)

An official UNSW Research Centre in the School of Mathematics & Statistics

formally established in 2021

Aim

Bring together UNSW’s full spectrum of data specialists to solve complex, real-world challenges, faced by governments and businesses

Data experts across UNSW

100+ data experts across UNSW’s broad and diverse faculties (Science, Arts, Engineering, Medicine, Law, Business, and Aus. Defence Force Academy)
All experts researching cutting edge applications using data science tools

We translate large volumes of data into knowledge to support decision-making.

uDASH

Our expertise:

Mathematics and Statistics
Machine Learning and Artificial Intelligence
Data Visualisation
Computational modelling and simulation
Non-linear dynamics and optimisation
Data privacy
Probablistic modelling
Risk quantification and management
Business, economics, and marketing
Spatial modelling
Big and complex data
Genomics and medical data
Ecological, environmental and climate data
Defence research

Workshop Program

Morning sessions 9:30am - 12:00pm

Interactive lecture style sessions 10:00am - 12:00pm
Special talks 9:30 - 10:00am

Lunch 12:00pm - 1:30pm

Free time to explore, get food, etc.
Optional but recommended fun activities
- School of Mathematics and Statistics outreach workshops (45min)
- Datasoc campus tour (30min)

Afternoon sessions 1:30pm - 4:00pm

Independent project work
groups of 5-6 students
Goal: visualise patterns, postulate hypothesis, statistical analysis and discussion
Demonstrators and instructors will give advice and recommendations

Monday

Morning session:

Induction
Introduction to data science
Software + Intro to programming with R

Afternoon session:

Intro to datasets
Project group selection

Tuesday

Morning session:

Meet a data scientist
- Nick Lillywhite, BioScout
Exploratory data analysis

Afternoon session:

Afternoon session: Work on projects (data exploration)

Wednesday

Morning session:

Meet a data scientist
- Peter Hartmann, Westpac Group
Statistical Modelling

Afternoon session:

Work on projects (statistical modelling)

Thursday

Morning session:

Datasoc: who they are, how can they improve your student experience
Data visualisation

Afternoon session:

Work on projects (visualising results and wrap up)

Friday

Presentations
Program wrap-up

Your instructors

Dr. Steefan Contractor

Dr. Ziyang Lyu

Dr. José Ferrer

Dr. Maeve McGillycuddy

Dr. Sean Gardiner

Dr. Peng Zhong

Dr. Boris Beranger

Dr. Prosha Rahman

Dr. Anikó Tóth

Contacts and Socials

For urgent matters find my on (Slack)

BREAK

Quick stretch, walk around, switch tables

What are you most excited for during the coming week?

I go on a hike, and everytime I spot a cockatoo, I note down the temperature and atmospheric pressure. Can I use this data to investigate the relationship between temperature and pressure?

What is data science?

Survey

Get inspired!

Go to //historyofdatascience.com and browse some profiles…
Or scroll through the timeline
Who is your favourite data science icon?
- an 18th century pioneer?
- a data revolutionary? a data hero?
- an artificial intelligence Jedi?

The four paradigms of research

Paradigm 1: Experimentation

Father of Modern science

Paradigm 2: Theory led experimentation

Paradigm 3: Numerical modelling

Paradigm 4: Data-intensive scientific discovery

The concept of these four paradigms of research was coined by Jim Gray, a 1998 Turing Award winner, in 2007.

Data science is more than just prediction

Introduction to programming with R and Rstudio

GPCC_colloquium_2023

Fri, 07 Jul 2023 23:47:42 +1000

Rainfall Estimates on a Gridded Network (REGEN)

Steefan Contractor

Contractor, S., Donat, M. G., Alexander, L. V., Ziese, M., Meyer-Christoffer, A., Schneider, U., Rustemeier, E., Becker, A., Durre, I., and Vose, R. S.: Rainfall Estimates on a Gridded Network (REGEN) – a global land-based gridded dataset of daily precipitation from 1950 to 2016, Hydrol. Earth Syst. Sci., 24, 919–943

Basic description

Daily estimates over 1950 - 2016
Gridded 1 degree latitude x 1 degree longitude resolution
Global land coverage

Purpose

Purpose built for climate studies with a long temporal record and consistent global spatial analysis
Based on a large in situ archive from combining GPCC with GHCN-Daily among others
Includes various statistical model error estimates
Also includes guidance for users less aware of issues with in situ based precipitation observations

In Situ Station Archive Description

Total stations: 135,178
Around 50K stations for each day
Min stations per day: 35,460
Max stations per day: 56,190

Component In Situ Archives

Three sources:

GPCC stations
GHCN-Daily stations
Collected during GEWEX workshops

Merging algorithm:

Lat + Lon match and
World Met. Org. (WMO) ID match or missing

Coordinates within 1º of each other and
WMO ID match or missing and
0.99 correlation between timeseries with 365 days of data of which at least 10d with >1mm precip

Quality Control Procedures

The automated QC procedures were identical to those applied to GHCN-Daily (Durre et al. 2010)
The procedure included two stages
Stage 1 does temporal checks
- multi-day accumulations
- duplicate data within timeseries
- frequent occurance of values
- world record exceedances
- outlier checks
- temporal consistency checks
Stage 2 does spatial checks
- checks whether values are consistent with negihbours

Durre, I., Menne, M. J., Gleason, B. E., Houston, T. G., and Vose, R. S.: Comprehensive automated quality assurance of daily sur- face observations, J. Appl. Meteorol. Clim., 49, 1615–1633

Interpolation Algorithm

Ordinary Block Kirging
Best Linear Unbiassed Estimator (BLUE)
Linear because the estimate is a weighted average of surrounding stations

$$\mathbf{Z}^*(s_0) = \sum_{i=0}^{N} λ_i\mathbf{Z}(S_i)$$

Best because we use the spatial structure (covariance) to determine the value of the weights
Unbiassed because the weights are constrained to add up to 1 and so the result cannot be biassed to any one station

$$\sum_{i=1}^N λ = 1$$

Ordinary Kriging assumes second order stationarity (mean and variance constant across domain)

$$\mathbf{Z}^*(s_0) = μ + ε(s_0)$$

Block implies that the algorithm produces gridded area-average estimates as opposed to point estimates

Two Flavours: All stations and Long Term Stations Only

The All stations based dataset interpolates all underlying stations
The Long Term version interpolates only stations with 40 complete years of data
A year is complete if all 12 months had at least 70% non-missing days

Uncertainty aware guidance for users

The uncertainty info includes Kriging Error (KE): a weighted average of modeled variance (between interpolation location and stations) and depends solely on the spatial distribution of stations and grid size, and
Yamamoto coefficient of variation (CV) (Yamamoto et al. 2000): weighted (by Kriging weights) average error between the estimate and the station values
Number of stations used for each grid estimate is also included

Yamamoto, J. K.: An Alternative Measure of the Reliability of Ordinary Kriging Estimates, Math. Geol., 32, 489–509

Quality Masks

A grid cell was left unmasked if:

It contained 60% of days in every decade with at least 1 station, and
both the KE and CV were under the 95th percentile (spatial distribution) of the temporally averaged (over 1950 - 2016) KE and CV respectively

Comparison with other global gridded datasets of monthly precipitation

Comparison with other global gridded datasets of daily precipitation

Comparison with regional daily precipitation datasets

Mean difference

SD of difference

Temporal correlation

Application: Global changes in precipitation

Trends in annual precipitation (1950 - 2016) (mm/yr)

Contractor, S., Donat, M. G., & Alexander, L. V. (2021). Changes in Observed Daily Precipitation over Global Land Areas since 1950. Journal of Climate, 34(1), 3–19.

Wet-day frequency changes between 1950-1983 and 1984-2016 (%)

Mean precipitation intensity changes between 1950-1983 and 1984-2016 (%)

Changes across the precipitation distribution between 1950-1983 and 1984-2016

Spatially, changes in precipitation seem complex, even stochastic at first
But a clear signal of positive precipitation changes in the high quantiles consitent with thermodynamic expectations is apparent
This signal dissappears for the most extreme precipitation again

Relative difference in area showing postive changes vs area showing negative changes

Synchronous changes between frequency and intensity

Mean changes in frequency and intensity are aligned in only around 1/3^rd of the grids
Extreme changes in frequency and intensity are aligned in almost 80% of areas globally

(My ideal) Future of climate datasets

All “observational” datasets are estimates from a statistical model consisting of aleatoric and epistemic uncertainties
If we stop thinking of observations as immutable facts and instead think of them as data generating models than we can ask more meaningful questions
E.g. for validation studies, instead of doing a grid cell by grid cell comparison we can calculate the conditional probability of the model output given the observations
To do this we need observations to be inherently probabilistic (the entire distribution), e.g. Risser et al. 2019
Artificial intelligence assisted inference can alleviate computational bottlenecks that traditionally made inference algorithms impractical in climate sciences, e.g. Zammit-Mangion et. al 2021 and Lenzi et. al 2023
As the examples demonstrate even a dataset of extremes is possible with this approach

Risser, M. D., Paciorek, C. J., Wehner, M. F., O’Brien, T. A., & Collins, W. D. (2019). A probabilistic gridded product for daily precipitation extremes over the United States. Climate Dynamics, 53(5), 2517–2538.

Zammit-Mangion, A., Ng, T. L. J., Vu, Q., & Filippone, M. (2021). Deep Compositional Spatial Models. Journal of the American Statistical Association, 0(0), 1–47.

Lenzi, A., Bessac, J., Rudi, J., & Stein, M. L. (2023). Neural networks for parameter estimation in intractable models. Computational Statistics & Data Analysis, 185, 107762.

Slides | Steefan Contractor

Aceas Multiage Seaice Classification

Bayesian updates to multi-age Antarctic sea-ice concentrations using GNSS-R data

Sea Ice

Remote Sensing of Sea Ice

GNSS-R

IUP Multiyear ice concentration and other sea ice types, Version AQ2 (Antarctic)

Sea Ice Signal in GNSS-R features?

Correlation amongst GNSS-R features and ice types

Geographic distribution

Ice concentrations as probabilities

Bayesian update

Which ML model?

Decision tree ensemble methods

LightGBM

Gaussian Mixture Discriminant Analysis

Robust Mixture Discriminant Analysis

Class rebalancing

SMOTE

Performance after class rebalancing

Updated ice probability density

UMAP (Uniform Manifold Approximation and Projection)

Performance after UMAP feature transformation

Updated ice probability density

Updated geographic distribution - winter

Updated geographic distribution - summer

Thank you

Deep Sea Talk

Observing oceans from above

A lot happens on the ocean surface

Sea Ice

Sea Ice

UK-Aus Space bridge Grant

Global Navigation Satellite System - Reflectometry

GNSS-R Features

water/ice and ice type classification

Thank you

Intro

Year 10 Data Science Work Experience Week

Acknowledgement of country

Health and Safety Induction

General safety in the classroom

Cold & Flu safety

Emergency Procedure

Emergency Evacuation

First Aid - Physical

First Aid - Mental

UNSW Year 10 Data Science Work Experience Week

Aim

Students will

UNSW Data Science Hub (uDASH)

uDASH

Workshop Program

Morning sessions 9:30am - 12:00pm

Lunch 12:00pm - 1:30pm

Afternoon sessions 1:30pm - 4:00pm

Monday

Tuesday

Wednesday

Thursday

Friday

Your instructors

Contacts and Socials

BREAK

What is data science?

Get inspired!

The four paradigms of research

Paradigm 1: Experimentation

Paradigm 2: Theory led experimentation

Paradigm 3: Numerical modelling

Paradigm 4: Data-intensive scientific discovery

Data science is more than just prediction

Introduction to programming with R and Rstudio

GPCC_colloquium_2023

Rainfall Estimates on a Gridded Network (REGEN)

Basic description

Purpose

In Situ Station Archive Description

Component In Situ Archives

Three sources: