1. CLUTTER MITIGATION DECISION (CMD) THEORY AND PROBLEM DIAGNOSIS
RADAR MONITORING WORKSHOP
ERAD 2010
SIBIU ROMANIA
Mike Dixon
National Center for Atmospheric Research Boulder, Colorado

2. Why do we need to know where clutter is in order to filter it effectively?
We need to filter normal-propagation (NP) clutter and anomalous-propagation (AP) clutter
The filters used remove weather power in some circumstances:
Velocity close to 0
Spectrum width close to 0
These conditions occur both in:
Clutter
Stratiform precipitation
We therefore need a technique which identifies likely locations of clutter
The filter is applied only at those gates with a high probability of clutter.

3. Problem - weather and clutter combined Reflectivity plot

4. Problem - weather and clutter combined Velocity plot

5. What happens if we filter everywhere?
Weather power removed
Filtered reflectivity – applying clutter filter everywhere

6. And if we use CMD?
Filtered reflectivity using CMD

7. Combined clutter and weather. Weather and clutter are distinct.

8. Combined clutter and weather spectrum
Combined clutter and weather spectrum. Weather and clutter overlap somewhat.

9. Combined clutter/weather spectrum
Combined clutter/weather spectrum. Weather and clutter overlap completely.

10. Notch filter – removes weather information

11. Adaptive filter – does not remove weather

12. Adaptive filter has a difficult time when clutter and weather overlap

13. Doppler Radar CMD

14. Motivation
Adaptive spectral clutter filters show great promise for intelligently filtering clutter power while leaving weather power largely unaffected.
However, these filters still remove weather power under the following circumstances:
the weather return has a velocity close to zero;
the weather return has a narrow spectrum width.
This tends to occur with stratiform weather in the region of the zero isodop.
In order to mitigate the problem, information other than that used by the filters must be used to determine whether clutter exists at a gate.

15. Principal feature fields for CMD
In order to identify gates with clutter, we use a number of so-called feature fields. These contain information which is independent of that used by the clutter filter.
The feature fields used in the latest CMD version are:
The TEXTURE of the reflectivity field – TDBZ.
The SPIN of the reflectivity field. This is a measure of how often the reflectivity gradient changes sign.
The Clutter Phase Alignment or CPA, which is a measure of the pulse-to-pulse stability of the returned signal.

16. TEXTURE of reflectivity - TDBZ
TDBZ is computed as the mean of the squared reflectivity difference between adjacent gates.
TDBZ is computed at each gate along the radial, with the computation centered on the gate of interest.
TDBZ at a gate is computed using the dBZ values for the 4 gates on either side of the gate of interest.
Computed for this gate
Using data from these gates

17. TDBZ feature field Example of TDBZ – Denver Front Range NEXRAD - KFTG

18. Reflectivity SPIN x y
For a point at which a gradient sign change occurs, let x be the reflectivity
change from the previous gate and y be the reflectivity change to the
next gate.
Then
SPIN CHANGE = (|x| + |y|) / 2

19. SPIN feature field
Example of SPIN – Denver Front Range NEXRAD – KFTG (SPIN is noisy in low SNR regions)
SPIN
DBZ

20. Clutter Phase Alignment - CPA
In clutter, the phase of each pulse in the time series for a particular gate is almost constant since the clutter does not move much and is at a constant distance from the radar.
In noise, the phase from pulse to pulse is random.
In weather, the phase from pulse to pulse will vary depending on the velocity of the targets within the illumination volume.

21. I,Q data is a complex number
power
phase I

22. CPA – theoretical phasor diagrams

23. CPA feature field
CPA is computed as the length of the cumulative phasor vector, divided by the sum of the power for each pulse.
CPA is computed at a single gate.
It is a normalized value, ranging from 0 to 1.
In clutter, CPA is typically above 0.9.
In weather, CPA is often close to 0, but increases in weather with a velocity close to 0 and a narrow spectrum width.
In noise, CPA is typically less than 0.05.
CPA was originally developed as a quality control field for clutter targets used for refractivity measurements.

24. Example of CPA – Denver Front Range NEXRAD - KFTG
CPA feature field
Example of CPA – Denver Front Range NEXRAD - KFTG
CPA
DBZ

25. Combining TDBZ, SPIN and CPA
The individual feature fields, TDBZ, SPIN and CPA, are combined into a single interest field using fuzzy logic.
First, each feature field is converted into an interest field, using a membership transfer function.
Interest fields have a range from 0.0 to 1.0.
The interest fields are assigned a weight.
The combined interest field is computed as a weighted mean of the individual interest fields.

26. Steps in computing the single-pol CMD
Compute TDBZ and TDBZ-interest
Compute SPIN and SPIN-interest
Compute CPA and CPA-interest
Step 2:
Compute Texture-interest = maximum of (TDBZ-interest, SPIN-interest)
Step 3:
Compute CMD value = fuzzy combination of CPA-interest and Texture-interest
Compute CMD flag: true if CMD >= 0.5, false if CMD < 0.5

27. Membership functions for single-pol CMD
-> (30,1)
-> (15,0)

28. Membership function combination as used in single-pol CMD
Weight=1.0
-> (30,1)
-> (15,0)
Weight=1.01

29. Creating combined interest field - CMD
TDBZ
SPIN
CMD
CPA

30. Logic for setting the clutter flag3.
CMD
> 0.5?
1. SNR > 3dB?
3. Set
Flag.

31. EXAMPLE OF APPLYING CMD

32. KFTG 2006/10/26, 1200 UTC
Reflectivity

33. VELOCITY, WIDTH
Radial velocity
Spectrum width

34. TDBZ, SPIN
TDBZ
SPIN

35. CPA, CMD
CPA
CMD

36. Clutter flag
CMD flag

37. Unfiltered reflectivity

38. Filtered reflectivity

39. DIAGNOSING ERRORS

40. THERE ARE 3 MAIN ERROR TYPES
FALSE DETECTIONS: the algorithm detects clutter incorrectly, so that the filter is applied excessively. This is particularly problematic when it occurs in the region of 0 velocity, since the filter cannot distinguish between clutter and weather.
MISSED DETECTIONS: the algorithm fails to identify clutter, and the filter is therefore not applied where it should be.
FILTER FAILURE: the CMD algorithm works correctly, but the filter fails to work effectively.

41. This is the most common error type.
FALSE DETECTIONS
This is the most common error type.

42. Example 1 - Filtered DBZ Filtered reflectivity using CMD
Filtered everywhere

43. Example 1 – VELOCITY and WIDTH
Radial velocity
Spectrum width

44. Example 2 : Filtered DBZ Filtered everywhere Filtered using CMD
Weather power removed
Filtered using CMD
Filtered everywhere

45. Example 2 : VEL, filtered VEL
Velocity
Filtered velocity

46. Example 3 : KFTG 2006/10/10, 1000 UTC Filtered everywhere
Weather power removed
Filtered everywhere
Filtered using CMD

47. Example 3 : VEL, filtered VEL
Velocity
Filtered velocity

48. Example 4 : KFTG 2006/10/09, 2200 UTC Filtered everywhere
Weather power removed
Filtered using CMD
Filtered everywhere

49. Example 4 : VEL, filtered VEL
Velocity
Filtered velocity

50. Example 4 : Operational NEXRAD
Filtered reflectivity
Filtered velocity

51. ERROR TYPE 2: MISSED DETECTIONS
It was found that in some clutter regions, CPA values can be lower than expected.

52. KEMX reflectivity 2009/04/22 21:22 UTC Also showing counties, interstates and US highways
We will investigate the region indicated by the ellipse

53. Unfiltered reflectivity

54. Filtered reflectivity
Ellipses highlight missed detections

55. Max interest for TDBZ/SPIN

56. Region 1 – CPA
This region has CPA values with considerable variability, with some values in the clutter region being as low as 0.2.

57. Characteristics of the CPA field
It was noted that in some of the clutter regions, there are a considerable number of gates with low CPA values. The CPA values can vary from high to low in adjacent gates.
Examining the phase time series for these gates shows that the phase can change significantly over a small part of the time series.
It is probable that 2 separate targets are illuminated during the dwell.
This phase change reduces CPA.
However, for the remainder of the time series the phase is stable.

58. Example - spectrum of clutter point with low CPA CPA = 0.19
A-scope X-axis: range (km)
Magenta line shows range
Red – unfiltered spectrum X-axis: samples
Pink – filtered spectrum
Power time series
X-axis: time
Phase time series
X-axis: time
Pulse-to-pulse phase
Difference time series
X-axis: time
Change in phase for part of the time series leads to low CPA value

59. ERROR TYPE 3: FILTER DOES NOT WORK EFFECTIVELY
The clutter filter can have problems dealing with multi-mode spectra.

60. Traffic clutter spectra
It appears that some of the missed detections and poor clutter filter performance are caused by traffic echoes.
The following slides show spectra from normal clutter echoes and suspected traffic echoes.
The traffic echoes exhibit multi-modal spectra. This makes them both difficult to detect as clutter, and difficult to filter with the current adaptive filters.

61. Unfiltered reflectivity
Note interstate 10 – shown in bold – traversing this area,
and smaller roads

62. Filtered reflectivity showing gates at which CMD and the clutter filter failed
Ellipse high-lights gates for which CMD and/or the clutter filter failed

63. Normal-propagation clutter signature CPA = 0.88
A-scope X-axis: range (km)
Magenta line shows range
Red – unfiltered spectrum X-axis: samples
Pink – filtered spectrum
Power time series
X-axis: time
Phase time series
X-axis: time
Pulse-to-pulse phase
difference time series
X-axis: time

64. Spectrum of suspected traffic targets CPA = 0.17
A-scope X-axis: range (km)
Magenta line shows range
Red – unfiltered spectrum X-axis: samples
Pink – filtered spectrum
Power time series
X-axis: time
Multi-modal spectrum
Phase time series
X-axis: time
Pulse-to-pulse phase
difference time series
X-axis: time

65. Spectrum of suspected traffic targets CPA = 0.30
A-scope X-axis: range (km)
Magenta line shows range
Red – unfiltered spectrum X-axis: samples
Pink – filtered spectrum
Power time series
X-axis: time
Multi-modal spectrum
Phase time series
X-axis: time
Pulse-to-pulse phase
difference time series
X-axis: time

66. THANK YOU