Enhancing high-resolution storm cell tracking

A multi-threshold TITAN algorithm in synergy with optical flow technique

Carlos Muoz *
Lopez , Li-Pen Wang and Patrick Willems
KU Leuven, BE (*carlos.munozlopez@bwk.kuleuven.be)

1. Introduction 3. Results & Discussion

TITAN (Thunderstorm Identification, Tracking, Analysis, and The performance of the enhanced TITAN tracking algorithm was
Nowcasting) is an object-based storm tracking algorithm, which assessed by measuring how well an one time-step deterministic
precipitation forecast (making use of the velocity field obtained
was developed for convective rain cell tracking (Dixon and Wiener,
from the tracking process) captures the real radar observations.
1993). However, due to the single-threshold setting, TITAN was The ROC (Receiver Operating Curve) curve analysis was used to
Single-Threshold Technique
not able to well identify adjacent storm cell evaluate the forecasting result, which allows the investigation of
cluster. In addition, its tracking algorithm the performance at a range of different rainfall intensities.
was developed based upon an overlapping
Relative Operating Characteristic (ROC )curve
technique, which has proven to be less Thresholds
(mm/hr)
effective for capturing fast moving storms 5 (True Positive) (False Positive)

and for high spatial-resolution rainfall data. 7.5

10
(False Negative) (True False)

15
20
27 Hit Rate: False Alarm Rate:
39

2. Methodology 48.5
55
65
=
+
=
+

75
ROC Curve: 100 Km x 100 Km
To improve the applicability of the original TITAN to high- 1
0.9

resolution rainfall data processing, two treatments have been 0.8

0.9
1

0.7

incorporated in this work.

0.8
0.6 0.7

H(Hit Rate)
For unskilled forecasts a ROC plot
0.6

H(Hit Rate)
0.5
0.5 52 dBZ ~ 75 mm/h
Multi-Threshold Technique 0.4 0.4 48 dBZ ~ 40 mm/h

1. Multi-threshold identification. will follow the diagonal line 0.3

0.2
0.3
0.2
0.1
42 dBZ ~ 15 mm/h

whereas an ideal forecast would 0.1

0

Based upon the hierarchical

0 0.05 0.1 0.15
F(False Hit Rate)
0

occupy the upper left triangle 0 0.2 0.4

F(False Hit Rate)
0.6 0.8 1

threshold segmentation (HTS)

entirely. ROC Curve: 50 Km x 50 Km

method (Peak and Tag, 1993), a 1

The performance of the proposed 0.9

multi-threshold technique was 0.8 1

treatments were evaluated at the

0.9
0.7 0.8

developed to better identify and 0.6 0.7

H(Hit Rate)

Single-Threshold Identification areas centred at Ghent urban 0.5

0.6
H(Hit Rate)

dBZ

isolate rainfall cells at small

0.5 52 dBZ ~ 75 mm/h
0.4 0.4 48 dBZ ~ 40 mm/h

catchment but with different 0.3

0.3 42 dBZ ~ 15 mm/h

scales. 0.2
0.2
0.1

C-Band Radar (~500 m/5 min) Resolution dBZ

extents (i.e. 20, 50 and 100 km). 0.1 0
0 0.05
F(False Hit Rate)
0.1 0.15

0
The results suggest that both 0 0.2 0.4
F(False Hit Rate)
0.6 0.8 1

treatments can lead to improved 1

ROC Curve: 20 Km x 20 Km

performance, and the combined 0.9

0.8
1

use of these two can result in the

0.9
0.7 0.8
0 50 Km
H(Hit Rate)

0.7
0.6

best performance, which is in

H(Hit Rate)

0.6
0.5 0.5 52 dBZ ~ 75 mm/h
0.4 0.4 48 dBZ ~ 40 mm/h

0 50 Km
Multi-Threshold Identification dBZ
particular evident for moderate 0.3
0.3
0.2
42 dBZ ~ 15 mm/h

0.2
and high rainfall intensities.
0.1

0.1 0
0 0.1 0.2 0.3 0.4
F(False Hit Rate)
0
0 0.2 0.4 0.6 0.8 1
F(False Hit Rate)
2. Optical Flow integration.
The optical flow technique (a 4. Conclusions and future work
field-based storm cell tracker)
was employed to provide an 0 50 Km
Results shows that the enhanced TITAN algorithm can better:
initial estimate of the storm cell Handle high-resolution and high intensity storm details.
movement. On the basis of this, a Tracking using Optical Flow dBZ Avoid faulty tracking between decaying and new born storm
combinatorial optimisation cells.
method based upon overlapping
Deal with mergers and splits.
and dynamic constraints
techniques, was developed in It also provides a great basis for future applications, where high-
order to match storm cells at two resolution rainfall estimates are critical; e.g., stochastic spatial-
successive time steps. 0 50 Km temporal rainfall modelling and nowcasting for urban hydrology.

Dixon, M. and Wiener, G. (1993), TITAN: Thunderstorm identification, tracking, analysis, and nowcastingA radar-based methodology, J. Atmos. Oceanic Technol., 10, 785797.
Peak, J. E. and Tag, P. M. (1994), Segmentation of satellite imagery using hierarchical thresholding and neural networks, J. Appl. Meteor., 33, 605616.
Wang, L.-P., et al., 2015. Enhancement of radar rainfall estimates for urban hydrology through optical flow temporal interpolation and Bayesian gauge-based adjustment, Journal of Hydrology. In press.