NDVI: how green
NDVI = (B08 - B04) / (B08 + B04)
- Near-infrared (B08) against red (B04) light.
- Higher means denser green vegetation; bare soil is low.
Method and limits
Method, limits and licences
How the numbers are made, what they cannot tell you and where the data comes from. Satellite data is updated every few days; clouds can hide a field for longer.
On this page
What it measures
Three numbers. REFARM computes two from Sentinel-2 for single fields; the third is an official indicator we only display.
NDMI: water in the leaves
NDMI = (B8A - B11) / (B8A + B11)
- Narrow near-infrared (B8A) against shortwave infrared (B11).
- Higher means more canopy moisture. It can fall before the field looks different.
CDI: drought class
Official EDO indicator
- Combines rain (SPI), soil moisture and vegetation (fAPAR) anomalies.
- Classes: no drought, watch, warning, alert, recovery, temporary soil-moisture and vegetation recovery.
- About 5 km grid, a new map every 10 days (a dekad). A municipality shows the most common class of the cells it touches.
Data sources
| Data | Used for | From |
|---|---|---|
| Sentinel-2 L2A | NDVI and NDMI of our Volos field, its 3D top surface and true-colour pictures | Copernicus Data Space Ecosystem, via Sentinel Hub |
| Combined Drought Indicator | Regional drought map and its history | European Drought Observatory (Copernicus Emergency Management Service) |
| Copernicus DEM GLO-30 | Terrain relief of our Volos field in 3D | Copernicus DEM public tiles, AWS Open Data registry |
| OpenStreetMap | Municipal boundaries and the base map | OpenStreetMap contributors |
Our Volos field
- The field on Our farms is our Volos field. Its boundary is provisional until the surveyed boundary is added.
- Everything measured on it is real: the Sentinel-2 values and Copernicus DEM heights for that boundary, run through the same steps as any field.
- Partner fields can be added once growers agree. Their data stays private unless they choose otherwise.
How the numbers are made
- Mask clouds. Keep only pixels the scene classification layer (SCL) marks as vegetation, bare soil or water.
- Use 10 m pixels. B04 and B08 are 10 m bands; B8A and B11 are 20 m bands, shown at 10 m.
- Shrink the field. Move the outline 5 m inwards so mixed pixels (tracks, hedges, neighbours) are not counted.
- Take field statistics. Store the mean NDVI and NDMI of the clear pixels per image date, with a quality value: the share of clear pixels.
- Build the normal. From the previous five years, per ten-day dekad: median, 10th and 90th percentile (p10, p90) and a spread. The charts draw p10 to p90 as the normal band.
- Compare with the normal. Each observation gets a difference and a robust z-score.
Normal and robust z-score in detail
- The year is split into 36 periods of about ten days (dekads). For each dekad and each of the previous five years we take the median of the good observations in that dekad and the dekads on either side.
- Across the years we keep the median, p10, p90 and a spread (1.4826 times the median absolute deviation).
difference = value - normal median
robust z = difference / max(spread, 0.04)
The spread never goes below 0.04, so a very steady normal does not turn small noise into a large score. A robust z of -2.5 means 2.5 spreads below the median.
Warning rules
One rule set for every field. A rule opens an event when its conditions hold on several good images in a row. Images with low quality, or without a reliable normal (under three years of data), are skipped.
Vegetation drop
NDVI clearly below normal on three good images in a row.
| Index | NDVI (Sentinel-2 L2A) |
|---|---|
| Counts only | Images with at least 0.6 of the field's pixels clear |
| Conditions (both) | robust z < -2.5 and difference < -0.12 |
| Opens after | 3 good images in a row meet the conditions |
| Severity | Warning. Critical if robust z < -4.0 |
| Closes after | 2 good images in a row that do not meet the conditions |
Moisture stress
NDMI clearly below normal on three good images in a row: water stress.
| Index | NDMI (Sentinel-2 L2A) |
|---|---|
| Counts only | Images with at least 0.6 of the field's pixels clear |
| Conditions (both) | robust z < -2.5 and difference < -0.08 |
| Opens after | 3 good images in a row meet the conditions |
| Severity | Warning. Critical if robust z < -4.0 |
| Closes after | 2 good images in a row that do not meet the conditions |
Data gap
No usable image for over 30 days, usually clouds.
| Counts only | NDVI images with at least 0.6 of the field's pixels clear |
|---|---|
| Condition | More than 30 days since the last good image |
| Severity | Information |
A warning is a prompt to look, not a diagnosis. Thresholds are written out by hand here; the version badge comes from the data shown. See the limitations.
Suggested tasks
Some warnings suggest one short check for a person. The grower accepts or ignores it; nothing is sent anywhere.
| Warning | Suggested task | Window |
|---|---|---|
| Vegetation drop | Walk the field, photograph the weakest area and record what you see. | 72 hours from when the task is suggested |
| Moisture stress | Check irrigation and soil moisture at 20 cm in the stressed part of the field. | 48 hours from when the task is suggested |
The check file
A suggested check can be saved as a JSON file, an open format other software can read. It holds the field and its outline, the action, the reason (event, type and severity), the time window, who could do it (a person or a machine, and what they need) and the status. Readers must ignore entries they do not know. A field on Our farms with a suggested check offers it as a download.
The 3D field view
Our Volos field as a small 3D model that turns slowly and can be dragged, built from the same data as the charts.
What it shows
- Outline: the field boundary, raised as a translucent block.
- Top surface: the latest usable Sentinel-2 NDVI map, greener where vegetation is denser, with its date.
- Relief: Copernicus DEM GLO-30 heights, cells of about 30 m, resampled to 64 by 64 and smoothed. Heights are stretched so gentle slopes show; the factor is in the model's text description.
What it cannot show
- Details under 30 m: ditches, banks, terraces, tree rows.
- The ground under the crop: the DEM is a surface model, so trees and buildings can raise it.
- More than one date: the top surface is one image and can be weeks old.
- Exact values: read the numbers in the field's charts and table.
Without WebGL 2 a still picture of the model is shown. With reduced motion on, it does not rotate by itself.
Limitations
Read these before relying on any figure on this site.
- Clouds and update rhythm
Sentinel-2 passes every few days and cannot see through clouds. The site is updated every few days, and a field can go weeks without a usable image. Every map and page shows its data date.
- 10 m pixels and small fields
Under about 0.3 ha, a field has only a few dozen pixels after the 5 m buffer, so one odd pixel moves its average. Fields under 0.1 ha are marked as small.
- The normal is short
It uses the previous five years only. One unusual year shifts it, and a dekad with under three years of data is not treated as reliable.
- Crop rotation can look like an anomaly
The normal belongs to the field, not the crop. A new crop or an early or late harvest differs from it without anything being wrong. The first vegetation rule raised many such events, so its thresholds were tightened.
- The cloud mask is not perfect
The SCL layer is automatic: it can miss thin cloud, haze and shadow, or mark clear ground as cloud. The cloud-probability option (CLD) does not flag cloud shadow at all. A low reading after a hazy day may be the atmosphere, not the crop.
- Drought map resolution and purpose
The CDI grid is about 5 km, so small municipalities can share a class, and a colour says little about one field. EDO provides it for information only. No usable value is shown as "No data".
- An early-stage service
The service is at an early stage. Rules and thresholds may change, and some pages or data may be missing. True-colour pictures are enlarged smoothly and hold no more detail than the 10 m data.
Data licences and credits
The datasets this site uses and the credit each licence asks for.
Sentinel-2 L2A
Credit
Contains modified Copernicus Sentinel data [year], processed by REFARM
Notes
Combined Drought Indicator (EDO)
Credit
European Drought Observatory, https://drought.emergency.copernicus.eu (C) European Commission - JRC, 2012 - 2026
Contains modified Copernicus Emergency Management Service information 2026
Reference and legal notice
Reference
Cammalleri et al. (2021) NHESS 21, 481–495
Legal notice
For information purposes only
Copernicus DEM GLO-30
Credit
produced using Copernicus WorldDEM-30 © DLR e.V. 2010-2014 and © Airbus Defence and Space GmbH 2014-2018 provided under COPERNICUS by the European Union and ESA; all rights reserved
Legal notice
The organisations in charge of the Copernicus programme by law or by delegation do not incur any liability for any use of the Copernicus WorldDEM-30
OpenStreetMap
Credit
© OpenStreetMap contributors
Software licences
Open-source software used
| Software | Licence |
|---|---|
| MapLibre GL JS | BSD-3-Clause |
| PMTiles | BSD-3-Clause |
| uPlot | MIT |
| Astro | MIT |
| Inter (typeface) | SIL Open Font License 1.1 |
What this is not
- Not agronomic advice: a warning says where to look first, not what to do.
- Not a diagnosis: the satellite sees colour and moisture signals, not the cause.
- Not an official drought bulletin: the map shows EDO data for information only.
- Not a continuous monitor: it is updated every few days, and clouds create gaps.
- Not endorsed by the European Union, the European Commission, ESA or any data provider.
Contact
Questions, corrections or ideas: info@refarm.online.