基于GOES影像波段组合的像素云覆盖得分计算及昼夜云覆盖状态识别方法咨询

Great question! For GOES-16 cloud detection using CMI_C13, CMI_C08, and CMI_C10, you’ve got two solid paths: interpretable heuristic threshold methods (perfect for quick GEE implementations) and more complex models if you need higher accuracy. Let’s break both down.

These rely on the physical properties of the GOES channels—no fancy training required. First, a quick recap of what each target band does:

• CMI_C08: 0.86μm near-infrared (NIR) channel—clouds reflect strongly here during the day, while most land surfaces have low reflectance (snow is an exception).
• CMI_C10: 10.3μm infrared window channel—measures brightness temperature (BT) of the surface or cloud top (colder = higher/cloudier).
• CMI_C13: 13.3μm infrared absorption channel—sensitive to upper-level moisture and cloud heights, which creates useful temperature differences with C10.

Core Heuristic Metrics to Combine #

1. Brightness Temperature Difference (BTD): C13 - C10

   • Negative BTD values indicate high, cold clouds (like cirrus) because C13 detects colder upper-atmospheric features than C10.
   • Positive or near-zero BTDs point to low clouds, fog, or clear land.

2. NIR Reflectance (CMI_C08)

   • Only useful during the day—clouds will have reflectance values > ~0.3, while most bare/vegetated land is < 0.2.

Day/Night-Specific Rules #

Since C08 is useless at night, split your logic:

• Daytime:

  // Cloud = high NIR reflectance AND (cold BT OR negative BTD)
  var dayCloud = c08.gt(0.3).and(c10.lt(270).or(btd13_10.lt(-2)));
  

• Nighttime:

  // Cloud = negative BTD (cirrus) OR very cold BT (high clouds) OR warm BT + positive BTD (low clouds/fog)
  var nightCloud = btd13_10.lt(-3).or(c10.lt(260)).or(c10.gt(280).and(btd13_10.gt(1)));
  

You can determine day/night using the GOES solar_zenith_angle band—day is when the angle is < 90 degrees. Here’s a full GEE snippet to tie it all together:

// Load GOES-16 MCMIPC dataset
var goes16 = ee.ImageCollection("NOAA/GOES/16/MCMIPC")
  .filterDate('2023-01-01', '2023-01-02')
  .first();

// Extract relevant bands
var c08 = goes16.select('CMI_C08');
var c10 = goes16.select('CMI_C10');
var c13 = goes16.select('CMI_C13');
var solarZenith = goes16.select('solar_zenith_angle');

// Calculate BTD (C13 - C10)
var btd13_10 = c13.subtract(c10).rename('BTD_13_10');

// Define day/night mask
var isDay = solarZenith.lt(90);

// Combine day/night cloud masks
var cloudMask = isDay.multiply(dayCloud).add(isDay.not().multiply(nightCloud)).rename('cloud_mask');
// 1 = cloud, 0 = clear land

If heuristics aren’t cutting it (e.g., you need to distinguish thin cirrus, fog, or snow from clouds), turn to machine learning. GEE has built-in tools for this:

Decision Trees/Random Forests #

These are great because they still retain some interpretability compared to deep learning. Steps to implement:

1. Gather labeled training data (use existing cloud products like MODIS MOD06, or manually label pixels in GEE).
2. Use features like your three bands, BTDs, solar zenith angle, and even other GOES channels if available.
3. Train a classifier and apply it to your GOES imagery.

Example snippet:

// Assume you have a labeled FeatureCollection with a 'label' field (0=clear, 1=cloud)
var classifier = ee.Classifier.smileRandomForest(15)
  .train(labeledData, 'label', ['CMI_C08', 'CMI_C10', 'CMI_C13', 'BTD_13_10', 'solar_zenith_angle']);

// Classify the image
var cloudMask = goes16.addBands(btd13_10).classify(classifier);

Neural Networks #

For the highest accuracy (especially with large datasets), you can use GEE’s ee.Classifier.smileFeedForward or export data to train a model externally (e.g., TensorFlow) and import it back into GEE. This is overkill for most basic cloud detection tasks though.

Start with the heuristic method first—it’s fast, easy to tweak, and works well for most general cases. If you notice consistent errors (e.g., misclassifying fog or snow), then invest in training a machine learning model.

内容的提问来源于stack exchange，提问作者Andrew