Dataset
Land-cover classifications and chainage-level vegetation metrics for islands of the middle Congo River derived from Sentinel-2 imagery for 2020, 2024 and 2025
Abstract
Land-cover classifications of vegetated islands along a ~400 km section of the middle Congo River, derived from cloud-free Sentinel-2 Level-2A Surface Reflectance Harmonized imagery for the dry seasons of 2020 (pre-drought), 2024 (drought) and 2025 (post-drought). The dataset accompanies dataset accompanies Steadman et al. (in submission), "Vegetation Loss and Response to Hydrological Drought on the Islands of the Congo River."
The deposit consists of three classified raster maps (GeoTIFF, 10 m resolution) and one tabular dataset (CSV):
- 2020_Landcover.tif, 2024_Landcover.tif, 2025_Landcover.tif: two-class supervised land-cover classifications based on a Random Forest classifier (500 trees) trained on Sentinel-2 bands B2, B3, B4, B8 and B11. Class codes: 1 = vegetated, 2 = non-vegetated.
- chainage_vegetation_metrics.csv: per-chainage summary table, with one row per 10 km chainage interval (260-640 km along the Congo River centreline upstream from Kinshasa, after Carr et al., 2019). Columns include percentage vegetated cover for each year, June-August 2024 mean precipitation (mm, from Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS)) and canopy-height composition (proportion of <1 m, 1-10 m, and >10 m vegetation derived from the ETH Global Sentinel-2 10 m Canopy Height dataset, 2020).
The study area covers a ~400 km corridor of the middle Congo River from just downstream of Mbandaka to a few kilometres upstream of Mpouya, spanning the border region between the Democratic Republic of the Congo and the Republic of the Congo. The corridor contains approximately 800 islands. Sentinel-2 imagery was mosaicked across MGRS tiles 33MXT, 33MXU, 33MYU, 33MYV and 33MZV.
Dry-season composites were generated from images acquired on 16 June, 1 July, 31 July and 8 August 2020; 20 June, 10 July, 25 July and 8 August 2024; and 13 May and 10 June 2025. The chainage CSV summarises conditions for these three years and additionally includes June-August 2024 cumulative precipitation derived from CHIRPS daily data.
Sentinel-2 Level-2A Surface Reflectance Harmonized imagery was accessed via Google Earth Engine. Cloud, cloud-shadow, cirrus and saturated pixels were masked using the Scene Classification Layer (SCL) with a 1-2 pixel buffer. Annual median composites were classified with a Random Forest model (500 trees, 70/30 train-test split) trained on manually digitised polygons of the four land-cover classes, using high-resolution historical imagery from Google Earth Pro and Bing Aerial Imagery as references. A majority filter was applied to smooth isolated misclassified pixels. Per-chainage metrics were derived in ArcGIS Pro using the Tabulate Area tool, with 10 km chainage polygons intersecting the basin-scale Congo River centreline (Carr et al., 2019) as zones. Percentage vegetated cover was computed as the vegetated area divided by the sum of vegetated and non-vegetated areas (water and sand classes were masked from the analysis).
The classifications and chainage metrics were produced to characterise the spatial extent and structural pattern of vegetation loss across Congo River islands during the extreme 2024 hydrological drought, and to test the hypothesis that vegetation loss is jointly controlled by hydroclimatic exposure (rainfall gradient) and vegetation structure.
Details
| Previous Info: |
No news update for this record
|
|---|---|
| Previously used record identifiers: |
No related previous identifiers.
|
| Access rules: |
Public data: access to these data is available to both registered and non-registered users.
Use of these data is covered by the following licence(s): http://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/ When using these data you must cite them correctly using the citation given on the CEDA Data Catalogue record. |
| Data lineage: |
Sentinel-2 Level-2A Surface Reflectance Harmonized scenes (COPERNICUS/S2_SR_HARMONIZED) were acquired via Google Earth Engine, filtered by cloud cover and date for the dry season of each target year, and mosaicked across MGRS tiles 33MXT, 33MXU, 33MYU, 33MYV and 33MZV. Cloud, shadow, cirrus and saturated pixels were masked using the Scene Classification Layer with a 1-2 pixel buffer, and annual dry-season median composites were derived. A Random Forest classifier (500 trees) was trained on manually digitised training polygons for four land-cover classes (vegetated, non-vegetated, water, sand) using bands B2, B3, B4, B8 and B11, with a 70/30 train-test split. Water and sand were then masked from the classification. Resulting rasters were aligned to a common 10 m grid, smoothed with a majority filter and exported as GeoTIFFs. Chainage-level metrics were generated in ArcGIS Pro using the Tabulate Area tool with 10 km chainage polygons (derived from the basin-scale Congo River centreline of Carr et al., 2019) as zones, applied to the classified rasters and the reclassified ETH Global Sentinel-2 10 m Canopy Height dataset (Lang et al., 2023). Per-chainage June-August 2024 rainfall was derived from CHIRPS daily precipitation aggregated to JJA totals and reduced over each chainage polygon. Data were produced by the project team and supplied for archiving at the Centre for Environmental Data Analysis (CEDA). |
| Data Quality: |
Data are as given by the data provider, no quality control has been performed by the Centre for Environmental Data Analysis (CEDA).
|
| File Format: |
These data are given in GeoTIFF and CSV format.
|
Related Documents
Process overview
Instrument/Platform pairings
| Sentinel 2 Multispectral Instrument (MSI) | Deployed on: Sentinel 2A |
Instrument/Platform pairings
| Sentinel 2 Multispectral Instrument (MSI) | Deployed on: Sentinel 2B |
Instrument/Platform pairings
| Sentinel 2 Multispectral Instrument (MSI) | Deployed on: Sentinel 2C |
Computation Element: 1
| Title | Sentinel-2 dry-season compositing and Random Forest land-cover classification in Google Earth Engine, with chainage-level summary statistics generated in ArcGIS Pro using the Tabulate Area tool. |
| Abstract | Land-cover mapping was performed in Google Earth Engine using ee.Classifier.smileRandomForest (500 trees) was trained on manually digitised polygons sampling four classes (vegetated, non-vegetated, water, sand) from Sentinel-2 L2A Surface Reflectance Harmonized median composites of June-August each target year. Cloud and shadow masking used the Scene Classification Layer with a 1-2 pixel buffer. Classification used bands B2, B3, B4, B8 and B11 with a 70/30 train-test split, validated using overall accuracy, Cohen's Kappa, Producer's Accuracy and User's Accuracy (Congalton, 1991). A majority filter was applied to reduce salt-and-pepper noise. Chainage-level metrics were generated in ArcGIS Pro using the Tabulate Area tool, with 10 km chainage polygons as zone features and the classified rasters / canopy-height raster as class rasters. The CHIRPS June-August 2024 cumulative rainfall raster was aggregated using Zonal Statistics (mean) over the same chainage polygons. |
| Input Description | None |
| Output Description | None |
| Software Reference | None |
| Output Description | None |
| Output Description | None |
| Output Description | None |
No variables found.
Temporal Range
2020-06-16T00:00:00
2025-06-10T23:59:59
Geographic Extent
0.0000° |
||
16.1180° |
18.2130° |
|
-2.6270° |