Measuring the Column Dependence of Read Noise in ACS/WFC Bias Frames

This study analyzes ACS/WFC bias frames to determine that read noise exhibits no column dependence but is influenced by row number due to accumulated dark current, while also finding that physical pre-scans have lower noise than science arrays, amplifiers A and C show an initial linear noise decrease, and masking hot pixels reduces column noise.

The Gist

Imagine the Hubble Space Telescope as a giant, high-end camera floating in space. Inside this camera is a special sensor called the ACS/WFC, which is essentially a massive digital grid of light-sensitive pixels (like the sensor in your smartphone, but much bigger and more sensitive).

This report is like a mechanic's logbook where two engineers, A.M. Guzman and M.C. McDonald, investigate a specific "glitch" or behavior in how this camera reads its own data.

Here is the story of their investigation, broken down into simple concepts:

1. The Setup: The "Reading" Race

Think of the camera sensor as a stadium filled with 4 million seats (pixels). To take a picture, the camera needs to read the data from every single seat. It doesn't read them all at once; it reads them in rows, like a runner passing a baton.

• The Problem: As the runner (the readout process) moves across the stadium, the seats get slightly warmer and start generating their own "ghost" signals (called dark current).
• The Result: The seats at the very end of the row (the last to be read) have accumulated more "ghost noise" than the seats at the start. This means the "noise" depends on the row number, not the column number.

2. The Mystery: Is There a "Column" Problem?

The engineers wanted to know: Does the noise also change depending on which column (vertical line) you are looking at?

Imagine the stadium has 4 different announcers (Amplifiers A, B, C, and D), each responsible for reading a specific section of the seats. The team checked the data from four different time periods (some recent, some from 2005) to see if any specific vertical line of seats was "noisier" than the others.

The Verdict: No.
Just like you wouldn't expect the noise to change based on which vertical line of seats you sit in, the data showed that read noise does not depend on the column number. It only depends on the row (how long the pixel has been waiting to be read).

3. The "Empty Hall" Surprise

The engineers noticed something odd at the very beginning of the data stream.

• The Science Area: This is the main stadium where the actual photos are taken. It has "ghost noise" because the pixels have been sitting there waiting to be read.
• The Pre-Scan Area: This is like an empty hallway before the stadium doors open. These pixels are read out immediately and haven't had time to accumulate any "ghost noise."

The Discovery: The "hallway" pixels (pre-scans) were much quieter (lower noise) than the stadium seats.
The Advice: When calculating the camera's noise level, ignore the hallway. Only measure the stadium seats, or your math will be wrong because you're mixing quiet empty space with noisy active space.

4. The "Slow Start" Anomaly

While looking at Amplifiers A and C, the engineers noticed a weird pattern at the very start of the reading process (the first few hundred columns).

• The Analogy: Imagine a runner who starts the race slightly out of breath, then settles into a steady pace.
• The Finding: Amplifiers A and C showed a tiny, linear drop in noise right at the beginning before leveling out. Amplifiers B and D didn't do this; they started steady.
• The Mystery: The engineers don't know why A and C act like they are "warming up," but they confirmed it happens.

5. The "Bad Apples" (Hot Pixels)

Finally, they looked at specific columns that were unusually noisy (like a row of seats that are all broken and buzzing). They wanted to know: Can we fix this by covering up the broken parts?

They tried two strategies:

1. Strategy A (The Trail): They looked for "trails" of bad pixels (like a line of people dropping trash) and tried to cover them up. Result: It didn't help much. The noise stayed high.
2. Strategy B (The Unstable Column): They looked for columns that were flagged as "unstable" (like a section of the stadium that is shaking). They covered up the entire unstable section. Result: Success! The noise dropped back to normal levels.

The Lesson: If a whole column is acting up because it's unstable, you have to mask (cover) the whole unstable section to get a clean reading. Just covering the obvious "trash" (hot pixel trails) isn't enough.

Summary: What Does This Mean for Hubble?

This report is a quality control manual. The engineers are telling the people who process Hubble images:

1. Don't worry about columns: The noise is consistent across vertical lines.
2. Ignore the pre-scans: They are too quiet and will mess up your noise calculations.
3. Watch out for "warming up" amplifiers: A and C have a slight dip in noise at the start.
4. Mask the unstable columns: If a column is flagged as unstable, cover it up to get a clean, accurate image.

By following these rules, the scientists can ensure that the beautiful images Hubble sends back to Earth are as clear and accurate as possible, free from the "static" of the camera's own electronics.

Technical Summary

Here is a detailed technical summary of the Instrument Science Report ACS 2023-03, "Measuring the Column Dependence of Read Noise in ACS/WFC Bias Frames."

1. Problem Statement

The Advanced Camera for Surveys (ACS) Wide Field Channel (WFC) on the Hubble Space Telescope operates in a Low Earth Orbit radiation environment. Over time, radiation damage has caused a linear increase in dark current, resulting in a growing population of "warm" and "hot" pixels.

• Read Noise vs. Row Number: It is well-established that read noise in ACS/WFC depends on the row number. This is due to "readout dark" (dark current accumulated during the ~100-second readout time), which increases linearly as pixels further from the amplifiers are read last.
• The Unknown: While row dependence is known, the column dependence of read noise had not been rigorously investigated. Specifically, it was unclear if read noise varies across columns, how it evolves over different anneal periods (thermal cycles used to reset the CCDs), and if specific columns exhibit elevated noise due to unstable hot pixels.

2. Methodology

The authors analyzed raw bias frames (zero-exposure images) from four distinct anneal periods to investigate these dependencies:

• Data Sources: Bias frames from four periods: 2022-08-17 (recent), 2022-03-28 (6 months prior), 2017-09-21 (6 years prior), and 2005-04-19 (pre-Servicing Mission 4).
• Noise Calculation:
  • Created difference images by subtracting subsequent bias files within an anneal period.
  • Applied sigma-clipping to mask cosmic rays (mimicking the acsrej step).
  • Calculated variance per column and derived read noise using the formula: $ReadNoise = \sqrt{Variance / 2}$.
  • Flipped the x-axis for Amplifiers B and D to align them with the readout direction of A and C.
• Data Quality (DQ) Analysis: Utilized DQ arrays from "superbias" and "superdark" reference files to identify flags for hot pixels (flag 64), warm pixels (flag 16), and unstable hot columns (flag 128).
• Statistical Analysis:
  • Piecewise Linear Fitting: Applied to trimmed data (excluding physical pre-scans) to detect linear trends in the initial columns of the science arrays.
  • Masking Experiments: Tested two strategies to reduce elevated noise in specific columns: (1) masking hot pixel trails found in superdark DQ arrays, and (2) masking unstable hot columns identified in superbias DQ arrays.

3. Key Contributions

• Confirmation of Column Independence: The study definitively establishes that read noise does not depend on column number, confirming that the primary driver of read noise variation is row number (readout dark accumulation).
• Pre-scan Characterization: Identified that physical pre-scan columns have significantly lower read noise (~0.5 $e^-$ lower) than science columns because they do not accumulate readout dark.
• Discovery of Initial Linear Decrease: Found a subtle, linear decrease in read noise within the first several hundred columns for Amplifiers A and C, a phenomenon not observed in Amplifiers B and D.
• Masking Efficacy: Demonstrated that masking unstable hot columns (even if not flagged in the DQ array for a specific epoch) effectively reduces elevated read noise values to average levels.

4. Key Results

A. General Trends

• No Column Dependence: Across all four anneal periods and all four amplifiers, there is no significant trend linking read noise to column number.
• Pre-scan Anomaly: Physical pre-scan columns consistently show lower average read noise. The authors recommend excluding these columns when calculating amplifier read noise for calibration to ensure accuracy.
• Evolution Over Time: For specific columns with unique characteristics (e.g., elevated noise in only one anneal period), there is no visual trend of read noise evolution over the duration of a single anneal period. The noise values remain relatively stable throughout the anneal cycle.

B. Amplifier-Specific Behaviors

• Amplifiers A & C: Both exhibit a slight linear decrease in read noise in the first ~350–600 columns.
  • Amplifier A: Break point between columns 350–450.
  • Amplifier C: Break point between columns 500–600.
  • Note: The cause of this behavior in A and C but not B and D remains unknown.
• Amplifiers B & D: Show no linear decrease or increase in the initial science columns; their read noise remains flat across the column range.

C. Elevated Noise Columns (Case Study: Column 873, Amplifier A)

• Hot Pixel Trails: Masking trails of hot pixels found in superdark DQ arrays (rows 397–400 and 1571–1575) failed to reduce the elevated read noise. This indicates hot pixel trails are not the primary cause of column-wide noise elevation.
• Unstable Hot Columns: Masking the entire unstable hot column (rows 0–1611) successfully reduced the read noise to average levels.
• DQ Flagging Discrepancy: The study noted that unstable hot columns are not always flagged as "unstable" (flag 128) in the DQ array for every anneal period. For example, Column 873 was flagged as unstable in the 2022-03-28 period but not in 2022-08-17, despite the presence of the hot trail in the Science (SCI) frame. This is because the flag depends on stability over the anneal period.

5. Significance and Recommendations

• Calibration Accuracy: The report provides critical guidance for the calacs pipeline. By excluding physical pre-scans from read noise calculations, calibration accuracy is improved.
• Reference File Creation: The study highlights that relying solely on DQ flags for unstable columns may be insufficient. Visual inspection or alternative masking strategies for elevated columns are necessary to prevent outliers from skewing reference file statistics.
• Instrument Health Monitoring: The identification of specific columns with persistent elevated noise (e.g., Col 873 on Amp A) aids in long-term monitoring of detector degradation and the effectiveness of annealing cycles.
• Future Analysis: The finding of a linear decrease in Amplifiers A and C suggests a need for further investigation into the specific readout electronics or charge transfer inefficiencies unique to those amplifiers.

In conclusion, this report refines the understanding of ACS/WFC noise characteristics, moving beyond the known row-dependence to characterize column behavior, and offers practical solutions for mitigating elevated noise in calibration data.