BDC Processing Flow

Several procedures are performed to generate BDC Data cube collections. Based on that, Figure 3 illustrates a diagram containing the main procedures used to generate BDC data products.

BDC Image collection ingestion and Data Cube generation

Figure 3 - BDC Image collection ingestion and Data Cube generation.

Image Collection builder

Atmospheric Correction

The atmospheric correction procedure relies on each sensor characteristics. Due to that, for each sensor different processes may be applied, for instance, atmospheric correction is performed for Sentinel-2/MSI images using the Sen2cor [9], while for Landsat-8/OLI this procedure is performed through LaSRC [21] and for CBERS-4/AWFI is executed by MS3 [4]. Once this process is performed and the resulting data is cataloged, Surface Reflectance collection are available to be used to produce Data Cube Collections.

Cloud Masking

Cloud masking algorithms can be used to detect undesirable areas, such as cloud, cloud shadows or snow. However, cloud masking algorithm also depends on the characteristic of each sensor. Based on that, for Sentinel-2 we use the standard SCL product provided by Sen2cor, for Landsat-8 the Fmask (version 4.2) is being used [14] while for CBERS-4 this processing is performed using CMASK due to it reduced number of spectral bands. Once this process is performed and the data is cataloged, the Surface Reflectance Image Collections becomes available.

Data Cube Builder

Warp (Merge, Reprojecting, Resampling and Griding)

In order to build the data cube collections, all input images must be at the same projection, using the same tile system and present the same spatial resolution. The Warp procedure perform this standardization and can be seen in Figure 4. Warp consists in cropping and mosaicking all images that superimpose a target tile of the common grid, for a specific date. This spatial mosaic is reprojected to the target tile reference system and all bands are resampled to a determined spatial resolution through a bilinear function, except for quality assessment band, which is resampled using nearest neighbor to avoid changes on the image values.

BDC Cube generation

Figure 4 - Data Cube generation.

Temporal Compositing

BDC Data Cube Collections can be categorized in two types, identity and composed. An identity data cube nature consists in using all available images from its sensors original acquisitions. Based on that, its temporal compositing function is identity. Temporal Compositing function can be used to generate regular series. This is performed by reducing the time dimension, which generates regularly spaced in time observations, here called composed data cubes.

Three temporal compositing functions are being used in BDC: ( i ) average, ( ii ) median, and ( iii ) stack. Considering the time dimension and a time step, e. g. monthly or 16 days, these temporal composition functions are applied on the identity data cubes, ideally on observations that are not detected as cloud or cloud shadow by the quality assessment band as considered in the following Table:

Cloud Mask

Nodata Value

Clear Data Values

Not Clear Data Values

Saturated Data Values

SCL

0

4, 5, 6, 7

2, 3, 8, 9, 10, 11

1

FMASK

255

0, 1

2, 3, 4

CMASK

0

127

255

The Average temporal composing consists in the average of the observed values. The Median temporal compositing consists in the median value of the observations. The stack temporal compositing consists in aggregating pixels from all images in the time interval according to each image quantity of valid pixels, e. g. a pixel from an image with efficacy of 95% (5% of cloud, cloud shadows or partial image) is more reliable of compositing the final image than a pixel from a 60% efficacy image. The mentioned compositing functions can be seen in Figure 5.

BDC Time Compositing Functions

Figure 5 - BDC Time Compositing Functions.

Note

If there are only not clear observations values for a time step, the stack temporal compositing outputs the value from the clearest image.

After the time compositing function is applied, the Brazil Data Cube Builder also calculates the NDVI and EVI spectral indices using the spectral bands of the time composed data cube. Besides that, for the time composed data cubes, three provenance bands are also generated for tracking characteristics of the composite period. The data product bands are CLEAROB, TOTALOB and PROVENANCE.