Explore satellite data.
Ten areas in Rondônia, Brazil. Sentinel-2 imagery from 2018 and 2024.
NDVI measures vegetation. The cutoff decides which patches count as dense vegetation: lower includes more; higher is stricter.
Loading satellite data…
Highlight changes to see where patches crossed the cutoff between 2018 and 2024. The 2018 image stays clear for comparison.
Possible vegetation lossAbove cutoff in both yearsPossible vegetation gainNot enough data
Select a point on either image to try the land classifier on a 640 m patch.
- Possible vegetation loss
- — of patches above the cutoff in 2018
- Area flagged for possible loss
- — usable area within those patches
Vegetation loss can indicate deforestation, but crops and dry conditions can also change NDVI. These results do not confirm tree loss.
How this estimate works
NDVI is calculated from near-infrared and red reflectance: (NIR − Red) / (NIR + Red). Each complete 64 × 64 pixel patch is compared using only pixels valid in both years. Patches with less than 90% shared valid data are excluded, along with incomplete patches at the image edges.
A patch is a forest proxy when its mean NDVI meets the selected threshold. Loss means it met the threshold in 2018 and fell below it in 2024. The percentage is lost patches divided by initial forest-proxy patches. Hectares count valid pixels in lost patches at 10 m resolution.
These are cloud-masked median composites from June 1 to September 30, with the end date excluded, in each year. Both images share the same grid and display colour stretch.
High NDVI also occurs in crops and other vegetation. These estimates have not been validated against an independent forest-loss dataset. Changing the threshold changes the estimate.
Classify a satellite patch.
Upload a small RGB satellite image, or select a patch from one of the ten comparisons.
Choose an image to begin. The model loads on the first classification.
Model prediction
—
ResNet50 trained on the ten EuroSAT classes. The model was trained on European imagery; Amazon classifications are experimental. Scores are model probabilities, not measured accuracy.
About the classifier
Your saved ResNet50 runs locally in the browser. Images are resized to 224 × 224, converted to RGB values between 0 and 1, then normalized using the same ImageNet channel means and standard deviations as the training notebook.
The ten labels are AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial, Pasture, PermanentCrop, Residential, River and SeaLake. A single patch produces one classification; it does not describe the composition of an entire large scene.
The forest-change tool uses NDVI rather than these land-class predictions.