Preprocessing

Turning raw satellite imagery into Analysis Ready Data

openEO exposes two flagship preprocessing workflows: Atmospheric Correction for optical sensors, and SAR Backscatter computation for radar sensors. Both are available in standard and CEOS CARD4L-compliant variants.

Atmospheric Correction

Sentinel Hub VITO EODC CARD4L

When a satellite captures light reflected from the Earth’s surface, that signal has passed through the atmosphere twice — once on the way down, once on the way back up. Scattering, absorption, and haze all distort what the sensor actually records. Atmospheric correction reverses those distortions, converting top-of-atmosphere (TOA) radiance into surface reflectance — the physical quantity that characterises what’s actually on the ground, independent of sun angle, viewing geometry, and atmospheric state.

Without atmospheric correction, multi-temporal comparisons are unreliable: a pixel that looks “brighter” in July might just be reflecting a clearer sky, not a healthier crop. Corrected data makes time series meaningful.

Standard process: atmospheric_correction()
CARD4L variant: ard_surface_reflectance()

Load L1C data first

Atmospheric correction must be applied directly to uncorrected L1C data. The preprocessing algorithm is tightly coupled to the raw sensor metadata — applying other processes (band maths, filtering) before correction may produce incorrect or unsupported results on some backends.

CARD4L: more than just correction

ard_surface_reflectance() follows CEOS CARD4L specifications and includes additional steps not available as separate processes — such as BRDF correction and standardised metadata output. Use this variant when interoperability and data product compliance matter.

Key parameters

Parameter	Description	Example values
`method`	The correction algorithm to apply. Backend-dependent.	`“iCor”`, `“SMAC”`, `“FORCE”`
`elevation_model`	DEM used for topographic correction (optional)	`“COPERNICUS_30”`
`atmospheric_correction_options`	Method-specific parameters passed as a dictionary	`{“adjacency_effect”: true}`

Supported backends & methods

Backend	Method(s)	Notes
VITO	`iCor`, `SMAC`	Tested with Sentinel-2. Angle bands must be selected manually.
EODC	`FORCE`	`ard_surface_reflectance()` via davidfrantz/force
Sentinel Hub	Backend default	Via VITO Geotrellis driver

Code examples

import openeo

connection = openeo.connect("https://openeo.vito.be").authenticate_oidc()

# Load raw L1C Sentinel-2 — include angle bands required by iCor
l1c = connection.load_collection(
    "SENTINEL2_L1C_SENTINELHUB",
    spatial_extent={
        "west": 3.758, "east": 4.088,
        "south": 51.292, "north": 51.393
    },
    temporal_extent=["2017-03-07", "2017-03-07"],
    bands=[
        "B04", "B03", "B02",          # RGB
        "B09", "B8A", "B11",          # SWIR / water vapour
        "sunAzimuthAngles", "sunZenithAngles",
        "viewAzimuthMean", "viewZenithMean"
    ]
)

# Apply atmospheric correction with iCor
l2a = l1c.atmospheric_correction(method="iCor")

l2a.download("sentinel2_icor.tif", format="GTiff")

library(openeo)

con <- connect("https://openeo.vito.be")
login()

p <- processes()

# Load raw L1C collection with required angle bands
l1c <- p$load_collection(
  id = "SENTINEL2_L1C_SENTINELHUB",
  spatial_extent = list(west=3.758, east=4.088, south=51.292, north=51.393),
  temporal_extent = c("2017-03-07", "2017-03-07"),
  bands = c(
    "B04", "B03", "B02",
    "B09", "B8A", "B11",
    "sunAzimuthAngles", "sunZenithAngles",
    "viewAzimuthMean", "viewZenithMean"
  )
)

# Apply iCor atmospheric correction
l2a <- p$atmospheric_correction(data = l1c, method = "iCor")

compute_result(graph = l2a, format = "GTiff", output_file = "sentinel2_icor.tif")

import { Connection, authenticate } from "@openeo/js-client";

const con = await Connection.connect("https://openeo.vito.be");
await con.authenticateOIDC();

const builder = await con.buildProcess();

// Load raw L1C collection
const l1c = builder.load_collection(
  "SENTINEL2_L1C_SENTINELHUB",
  { west: 3.758, east: 4.088, south: 51.292, north: 51.393 },
  ["2017-03-07", "2017-03-07"],
  ["B04","B03","B02","B09","B8A","B11",
   "sunAzimuthAngles","sunZenithAngles","viewAzimuthMean","viewZenithMean"]
);

// Apply atmospheric correction
const l2a = builder.atmospheric_correction(l1c, { method: "iCor" });

const result = builder.save_result(l2a, "GTiff");
await con.computeResult(result, { filename: "sentinel2_icor.tif" });

SAR Backscatter

Sentinel Hub EODC Orfeo Toolbox CARD4L experimental

Synthetic Aperture Radar sensors like Sentinel-1 see through clouds and work day and night — but the raw data they produce (GRD, SLC) is not directly interpretable as a surface measurement. It must be calibrated, terrain-corrected, and normalised before it can be used analytically. This pipeline is collectively called backscatter computation.

The process converts the raw radar signal into a physically meaningful quantity — typically sigma0 (surface scattering cross-section) or gamma0 (terrain-flattened reflectivity) — in either linear or decibel scale. Terrain flattening is essential for hilly regions: without it, slopes facing the sensor appear artificially brighter than flat terrain, corrupting any analysis based on backscatter magnitude.

Standard process: sar_backscatter()
CARD4L variant: ard_normalized_radar_backscatter()

Apply before other processes

Like atmospheric correction, SAR backscatter processes are tightly coupled to raw sensor metadata. Only spatial and temporal filters (filter_bbox, filter_temporal) should be applied before calling sar_backscatter(). Applying band maths or other transformations first will fail or produce incorrect results.

Choosing sigma0 vs gamma0

Use gamma0 (terrain-flattened) when comparing backscatter across varying terrain — it removes slope-induced radiometric bias. Use sigma0 for flat terrain or when CARD4L compliance is not required. The ard_normalized_radar_backscatter() process always produces gamma0 by CARD4L specification.

Key parameters

Parameter	Description	Example values
`coefficient`	The radiometric quantity to compute	`“gamma0-terrain”`, `“sigma0-ellipsoid”`, `“beta0”`
`elevation_model`	DEM used for terrain correction (optional —
backend default otherwise)	`“COPERNICUS_30”`, `“MAPZEN”`
`noise_removal`	Apply thermal noise removal (recommended)	`true` / `false`
`options`	Backend-specific processing options as a dictionary	`{“orthorectify”: true}`

Supported backends & implementations

Backend	Process(es)	Implementation	Notes
Sentinel Hub via VITO	`sar_backscatter`, `ard_normalized_radar_backscatter`	Sentinel Hub SAR	Full CARD4L support
EODC	`sar_backscatter`	Sentinel-1 Toolbox	Sentinel-1 GRD only
Orfeo Toolbox via VITO	`sar_backscatter`	openeo-geopyspark-driver	sigma0 only; not CARD4L compliant

Code examples

import openeo

connection = openeo.connect("https://openeo.vito.be").authenticate_oidc()

# Load Sentinel-1 GRD — spatial/temporal filters are fine before backscatter
s1grd = (
    connection.load_collection(
        "SENTINEL1_GRD",
        bands=["VH", "VV"]
    )
    .filter_bbox(west=2.590, east=2.895, north=51.221, south=51.069)
    .filter_temporal(extent=["2019-10-10", "2019-10-10"])
)

# CARD4L-compliant normalised radar backscatter (gamma0, terrain-flattened)
nrb = s1grd.ard_normalized_radar_backscatter()

# Submit as batch job and download results
job = nrb.execute_batch()
job.get_results().download_files("./output/")

Standard sar_backscatter() with explicit parameters

# More control: use sar_backscatter() directly
backscatter = s1grd.sar_backscatter(
    coefficient="gamma0-terrain",
    elevation_model="COPERNICUS_30",
    noise_removal=True
)
job = backscatter.execute_batch(
    title="S1 Backscatter Test",
    description="gamma0 terrain-flattened over Belgium"
)

library(openeo)

con <- connect("https://openeo.vito.be")
login()

p <- processes()

# Load Sentinel-1 GRD with spatial and temporal filters
s1grd <- p$load_collection(
  id = "SENTINEL1_GRD",
  bands = c("VH", "VV")
)

s1grd <- p$filter_bbox(
  data = s1grd,
  extent = list(west=2.590, east=2.895, south=51.069, north=51.221)
)

s1grd <- p$filter_temporal(
  data = s1grd,
  extent = c("2019-10-10", "2019-10-10")
)

# Apply CARD4L-compliant normalised radar backscatter
nrb <- p$ard_normalized_radar_backscatter(data = s1grd)

# Submit batch job
job <- create_job(
  graph = nrb,
  title = "S1 CARD4L NRB"
)
start_job(job)

# Monitor and download when complete
job_status <- describe_job(job)
download_results(job, folder = "./output/")

import { Connection } from "@openeo/js-client";

const con = await Connection.connect("https://openeo.vito.be");
await con.authenticateOIDC();

const builder = await con.buildProcess();

// Load Sentinel-1 GRD
let s1grd = builder.load_collection(
  "SENTINEL1_GRD",
  null, null,
  ["VH", "VV"]
);

// Apply spatial and temporal filters first
s1grd = builder.filter_bbox(s1grd, {
  west: 2.590, east: 2.895,
  south: 51.069, north: 51.221
});

s1grd = builder.filter_temporal(s1grd, ["2019-10-10", "2019-10-10"]);

// CARD4L normalised radar backscatter
const nrb = builder.ard_normalized_radar_backscatter(s1grd);

const result = builder.save_result(nrb, "GTiff");

// Submit as batch job
const job = await con.createJob(result, { title: "S1 NRB Belgium" });
await job.startJob();

// Poll until done, then download
const results = await job.getResults();
for (const asset of Object.values(results.assets)) {
  console.log("Download:", asset.href);
}

Performance considerations

Both atmospheric_correction and sar_backscatter are among the most computationally expensive operations in openEO. Before including them in your workflow, consider:

Use batch jobs — synchronous execution will time out for anything beyond a small test area.
Filter early — always apply filter_bbox and filter_temporal before calling these processes to minimise the data volume processed.
Pre-corrected collections — many openEO backends offer analysis-ready collections (e.g. Sentinel-2 L2A, Sentinel-1 backscatter products) that have already been corrected. Check your backend’s collection catalogue before computing from raw data.
Cost awareness — backends may charge more for preprocessing-heavy workflows. Profile on a small tile before scaling up.

When to skip preprocessing

If your backend offers a collection like SENTINEL2_L2A or SENTINEL1_BACKSCATTER, use it directly — it is the output of these processes, pre-computed and stored at much lower retrieval cost.

Related processes

Process	Description
`load_collection()`	Entry point — load a raw L1C or GRD collection
`filter_bbox()`	Spatial filter — always apply before preprocessing
`filter_temporal()`	Temporal filter — always apply before preprocessing
`mask_scl_dilation()`	Cloud masking on corrected optical data
`save_result()`	Export your processed cube to file

Next: Spatial Operations →