3. What we do in Transight
Cloud based data
solution deployment
Enterprise scale
data platform with
Pnda
Customized ML
modeling and Apps
4.
5. Problem
➢ The problem exists not just in developing countries, but also in
developed countries.
➢ Huge private and public property loss during the flood disaster
➢ Life safety of the civilian live in the city
Resource: Smith, B., & Rodriguez, S. (2017). Spatial Analysis of High-Resolution Radar Rainfall
and Citizen-Reported Flash Flood Data in Ultra-Urban New York City. Water, 9(10), 736.
16. Challenges of this project
Get the volume of rainfall for each flood station
Flood data collected irregularly
Predict the flood severity of next 3 hours
17. Challenges and mitigations (1/3)
Calculate the volume of rainfall for each flood station based on distance
Distance-based weighted rainfall
18. Challenges and mitigations (2/3)
Flood data collected irregularly - frequency based on flood level
Split data by a fixed time window (1 hour)
0m 10m 20m 30m 40m 50m
0
0 7 4
6 22 21 20 21
3 6 3 1 0
0 11 10 10 9
9 9 8 7 5 4
19. Challenges and mitigations (2/3)
Flood data collected irregularly - frequency based on flood level
Split data by a fixed time window (1 hour)
0m 10m 20m 30m 40m 50m
0
0 7 4
6 22 21 20 21
3 6 3 1 0
0 11 10 10 9
9 9 8 7 5 4
27. Data enrichment
● Google Maps API
○ Elevations: elevation, west, east, north, south
● Gaode Maps API
○ Geo-location profiling: restaurants, tunnels,
metro stations and etc.
South
North
EastWest
29. Seq2Seq model in a nutshell (1/3)
● Seq2Seq models: recurrent neural networks for sequence to sequence
prediction
● Why?
○ Allow us to process information that has a time or order element to it
○ Preserve information that couldn’t be done via normal neural network
Sentence: [‘I’, ‘am’, ‘a’, ‘student’, ‘.’]
Bitcoin prices: [8191., 8267., 8251., 8261., 8240., 8219., 8212., 8212., 8206., 8155.]
30. Seq2Seq model in a nutshell (2/3)
● Use cases: neural machine translation, chatbots, auto reply, time series and
more
Luong, et al. "Neural machine translation (seq2seq) tutorial." (2017): 1532-1543.
31. Seq2Seq in a nutshell (3/3)
● Vanilla RNN, LSTM or GRU
● Single-layer or multiple-layer
LSTM
33. Define the performance metrics
Matching the prediction goal with the performance metrics
● The higher the flood level, the more crucial to predict it correctly
⇒ customized mae: mae_10, mae_20
37. LSTM LSTM LSTM
h1 h2
FC FC
LSTM LSTM LSTM
h1 h2
Encoder Decoder
Seq2Seq Baseline
Start
End
X1 X2 Xt
38. LSTM LSTM LSTM
h1 h2
FC FC
LSTM LSTM LSTM
h1 h2
Q1 Q3
Encoder Decoder
Seq2Seq + weather forecast
Concat Concat
Start
End
X1 X2 Xt
39. FC
LSTM LSTM LSTM
FC FC
LSTM LSTM LSTM
Concat
FC
Concat
FC
Concat
Encoder Decoder
Seq2Seq + weather + Geoinfo
Q1
Concat
Q3
Concat
Start
End
Q1 Q3
X1 X2 XtS1 S2 S2
40. Results
● Performance of the above three models
Model rmse mae_10 mae_20
Seq2Seq baseline 0.020 3.783 3.258
Seq2seq + weather
forecast
0.021 3.411 3.005
Seq2Seq + weather
+ geoinfo
0.022 3.322 2.941
Overall performance of the model has reached 60%-80% in
different levels of flood level prediction (20cm to 50cm+)
41. More about the models
● t =12
○ 12 hrs of history data used to predict the next 3 hours
● Hyperparameters
○ Optimizer: adam
○ epochs=50
43. Wrap up
Results: The model we deployed can predict the flood
level 3 hrs in advance with high reliability.
Action: Government is testing our prediction model
against their experts experience - to list flood severity 3
hrs in advance in the city.
Benefits: Once in production, city will have better flood
prevention and rescue system, save millions per year.