Let's say you have a Sentinel-2 satellite image mosaic over a large area, for example Russia. The image is supposed to represent the surface reflectance of the area. How do you know if the pixel values are actually correct and that the atmospheric correction and the radiometric normalization of the mosaic have been successful?

Having worked at Terramonitor Labs for a couple of years on this topic I know that there is no simple answer. Actually, we noticed that there did not exist a scientifically verified validation process for Sentinel-2 mosaics. We wanted to be sure that the satellite image products we provide to our customers have the best possible quality. That's why we have designed and implemented the world's first global benchmarking process for Sentinel-2 mosaics.

The basic idea of the validation process is fairly simple. We take a single cloudless satellite image, called the reference image, and compare it to the mosaic of the same area from the same time period. To avoid overfitting, the reference image is not used in the mosaicing process. In other words, we try to replicate the reference image using a series of other images of the same area. The error between the mosaic and the reference image is measured by calculating the pixelwise mean absolute percentage error (MAPE) of the mosaic for each band.

Figure 1. RGB visualizations of (a) a standard Sentinel-2 mosaic provided by another Sentinel-2 image provider (b) Terramonitor Sentinel-2 mosaic and (c) the reference image.

MAPE is used to describe the overall quality of a mosaic. Having said this, we acknowledge that the process has the following internal sources of error:

  1. The reference image does not represent the surface reflectance exactly
  2. Surface reflectance changes over time, even daily, especially in water areas

Despite these sources of error, MAPE seems to be a good indicator of quality.

The following example shows the error values of a mosaic produced by Terramonitor and a mosaic produced by another Sentinel-2 image provider. In both cases, MAPE and MdAPE (median absolute percentage error) are calculated by comparison to the same reference image in Riga, Latvia.

Table 1. Mean absolute percentage error (MAPE) and median absolute percentage error (MdAPE) for the standard Sentinel-2 mosaic and the Terramonitor Sentinel-2 mosaic. MAPE and MdAPE are calculated pixelwise for each of the R,G,B,NIR bands. The last column shows the average MAPE and MdAPE over the four bands.

The results in Table 1 show that the quality of the Terramonitor product is significantly better than the quality of the other product, in terms of the error measures MAPE and MdAPE. The result is also seen by looking at the RGB visualizations in Figure 1 and the error images in Figure 2.

Figure 2. Error maps of (a) the standard Sentinel-2 mosaic and (b) the Terramonitor Sentinel-2 mosaic. Blue represents low and yellow represents high error values. 

To ensure that our methods work in different geographical areas we have defined a global reference dataset consisting of 17 Sentinel-2 tiles worldwide (see Figure 3). This dataset enables continuous validation of new improvements to the Sentinel-2 Mosaic Service.

Figure 3. Global reference dataset locations

Contact us to get your verified Sentinel-2 Mosaic today!


+358 50 360 8096

Figure 4. Validated Sentinel-2 Mosaic