在R中读取Sentinel-2的JP2波段获取异常高值，寻求与SNAP显示值匹配的转换系数

Great question—this is a super common confusion with Sentinel-2 L1C data, and it all boils down to data scaling between raw digital numbers (DN) and top-of-atmosphere (TOA) reflectance.

Here's the full breakdown: #

• What R is reading: The JP2 files from Sentinel-2 L1C are stored as 16-bit unsigned integers, which are raw DN values (typical range 0–65535). Your terra output showing max values around 18k is totally normal for non-cloudy, non-snowy scenes.
• What SNAP is showing: By default, SNAP automatically applies the official Sentinel-2 scaling coefficients to convert those raw DN values to TOA reflectance—a physical unit that ranges roughly 0–1 (though bright targets like clouds can push it slightly higher). That’s why you see values around 0–0.15.

The conversion formula you need #

For Sentinel-2 L1C products, the official conversion is straightforward:

toa_reflectance = dn_value * 0.0001 + 0

The scale factor (0.0001) and offset (0) are defined in your scene's metadata file (MTD_MSIL1C.xml).

To apply this in R with terra:

# Convert your SpatRaster to TOA reflectance
r_toa <- r_10 * 0.0001

# Check the new range (should match SNAP's ~0–0.15)
minmax(r_toa)

Verify with your data #

Your max DN value is 18815—multiplying by 0.0001 gives 0.18815, which lines up closely with the ~0.15 range you saw in SNAP. The slight difference is likely because SNAP was showing the range of a scene subset, or excluding extreme bright pixels like thin clouds.

Bonus: Use metadata for robust conversion #

Instead of hardcoding 0.0001, you can pull scaling values directly from the scene's metadata to make your code flexible (in case of future product updates):

library(XML)

# Path to your scene's MTD_MSIL1C.xml file
meta_path <- "path/to/your/MTD_MSIL1C.xml"
meta_doc <- xmlParse(meta_path)

# Extract official scale factor and offset
scale_factor <- as.numeric(xmlValue(getNodeSet(meta_doc, "//Scaling_Factor")[[1]]))
offset <- as.numeric(xmlValue(getNodeSet(meta_doc, "//Offset")[[1]]))

# Apply the conversion
r_toa <- r_10 * scale_factor + offset

Why SNAP export is so slow #

SNAP does more than just basic scaling when you export—depending on your settings, it might be running additional processing (like resampling, atmospheric correction, or format conversion) on the full 10980x10980 scene, which is computationally heavy. Doing the scaling directly in R is way faster and more efficient.

内容的提问来源于stack exchange，提问作者candelas762