For the past few weeks, we've been looking into the parametric controls on change signals in datasets downscaled using GARD. @gutman asked that I put a little writeup here to document what we've looked at to date and allow for discussion on the next steps.
First, the figure that kicked off this inquiry:
![pasted_image_at_2017_10_24_12_07_pm_720](https://user-images.githubusercontent.com/2443309/33281873-91e5e9f4-d35b-11e7-8118-7d51aca37dad.png)
Caption: Change in precipitation (2080s-1980s) for 7 different downscaling methods. The WRF-50km dataset was used to drive the four GARD methods. GARD tends to show a stronger drying trend than LOCA, BCSD, or WRF-50km, particularly when circulation variables (PR2, PR3) are used.
Exploratory Analysis
Evaluation of Raw GARD output (no random errors, no QM)
We have looked at a series of GARD experiments using different predictor variables and transforms:
Cube Root (GARD_downscaling_production_20170929)
- PassThrough: PREC_TOT
- PR, PA, AR:
- 1: PREC_ACC_NC, PREC_ACC_C
- 2: PREC_ACC_NC, PREC_ACC_C, U, V, PSFC
- 3: U, V, PSFC
Cube Root (GARD_downscaling_production_20171017)
Cube Root (GARD_downscaling_production_20171103):
- PR, PA, AR:
- 1c: PREC_TOT
- 2a: PREC_TOT, U, V
- 2b: PREC_TOT, U, V, QVAPOR
- 2c: U, V, QVAPOR
Cube Root (GARD_downscaling_production_20171116):
- PA:
- 1d: PREC_TOT
- 2d: PREC_TOT, U, V
Log Root (GARD_downscaling_production_20171103b):
- PR, PA, AR:
- 1c: PREC_TOT
- 2c: PREC_TOT, U, V
5th Root (GARD_downscaling_production_20171103c):
- PR, PA, AR:
- 1c: PREC_TOT
- 2c: PREC_TOT, U, V
Evaluation of GARD output after random errors
Cube Root (GARD_downscaling_production_20170929)
- PassThrough: PREC_TOT
- PR, PA, AR:
- 1: PREC_ACC_NC, PREC_ACC_C
- 2: PREC_ACC_NC, PREC_ACC_C, U, V, PSFC
- 3: U, V, PSFC
Cube Root (GARD_downscaling_production_20171017)
Evaluation of GARD output after quantile mapping
Cube Root (GARD_downscaling_production_20170929)
- PassThrough: PREC_TOT
- PR, PA, AR:
- 1: PREC_ACC_NC, PREC_ACC_C
- 2: PREC_ACC_NC, PREC_ACC_C, U, V, PSFC
- 3: U, V, PSFC
Cube Root (GARD_downscaling_production_20171017)
Evaluation of trends/changes in predictor variables
I looked at the following WRF-50km variables: PREC_ACC_NC, PREC_ACC_C, U, V, PSFC, T2
Single point experiments
This work is currently underway. Exploring behavior of GARD for how analog sampling is done, transforms, and different input variables.