README.md

---
license: apache-2.0
pretty_name: OlmoEarth Land Cover Change (LCC) Dataset
tags:
- remote-sensing
- earth-observation
- change-detection
- sentinel-2
size_categories:
- 1K<n<10K
---

# OlmoEarth Land Cover Change (LCC) Dataset

This dataset contains point-based annotations of land cover change, used to
train the OlmoEarth LCC model (https://olmoearth-lcc.allen.ai). The model
detects recent land cover change from Sentinel-2 time series: it inputs a Sentinel-2
image time series with 16 quarterly images (to establish a historical baseline) and 4
recent biweekly images (to detect changes soon after they occur) and predicts, per pixel,
whether a land cover change occurred at some point during the time series, and if so,
the source and destination land cover categories, the start and end dates of the change,
and a fine-grained change category.

We also provide GeoTIFF outputs produced by running the trained model at scale. See
[Model output summary rasters](#model-output-summary-rasters) below.

The dataset is intended to be paired with Sentinel-2 images, but can be used with other
modalities as long as they are temporally aligned. We do not include the image inputs
here since they can be obtained based on the timestamps in the provided annotations.

## Annotation format

Annotations are provided as JSON files. Each file is a list of entries, where
each entry corresponds to one 128x128-pixel spatial window at 10 m/pixel in a
UTM projection:

```json
{
  "projection": {"crs": "EPSG:32651", "x_resolution": 10, "y_resolution": -10},
  "bounds": [27130, -161561, 27258, -161433],
  "window_name": "example_window",
  "group": "default",
  "time_range": ["2017-01-01T00:00:00+00:00", "2024-01-01T00:00:00+00:00"],
  "positive_points": [
    {
      "lon": 121.5,
      "lat": 14.6,
      "pre_change": "2020-01-15",
      "first_date_change_noticeable": "2020-04-27",
      "post_change": "2020-07-15",
      "pre_category": "tree",
      "post_category": "urban/built-up",
      "post_change_category": "new_building"
    }
  ],
  "negative_points": [
    {"lon": 121.51, "lat": 14.61}
  ]
}
```

Entry fields:

- `projection` / `bounds`: the UTM projection and pixel bounds
  `[min_col, min_row, max_col, max_row]` of the window. Pixel coordinates are
  projection coordinates divided by the resolution (10 m/pixel; `y_resolution`
  is negative so row indices increase southward).
- `window_name` / `group`: identifiers for the entry.
- `time_range`: the period over which the annotations are valid. In particular,
  negative points are only confirmed as no-change within this time range.
- `positive_points`: points where a land cover change was verified.
- `negative_points`: points verified as having no change. These only have
  `lon` and `lat`.

### Positive point fields

Each positive point has a location (`lon`, `lat` in WGS84), three dates, and
two land cover categories:

- `pre_change`: the last date at which the location still appears in its
  pre-change state in the imagery.
- `first_date_change_noticeable`: the first date at which the change starts to
  become noticeable.
- `post_change`: the first date at which the change appears complete.
- `pre_category`: the land cover category before the change.
- `post_category`: the land cover category after the change.

The land cover categories (used for both `pre_category` and `post_category`)
are:

`bare`, `burnt`, `crops`, `fallow/shifting cultivation`, `grassland`,
`Lichen and moss`, `shrub`, `snow and ice`, `tree`, `urban/built-up`, `water`,
`wetland (herbaceous)`.

In addition to the source/destination land cover categories, more recently
annotated positive points carry fine-grained change-category fields. These
fields were introduced partway through annotation, so **some positive points
have them and others do not**: a point either has at least one of the three
fields set, or none of them. During training, the change-category heads are
only supervised at points that have at least one of these fields; for such
points, any unset sibling field is treated as an explicit "none" label, while
points without any of the fields are masked out for these heads.

The three fields describe different aspects of a change, and a point is
annotated with one `pre_change_category` and/or one `post_change_category`,
*or* one `same_change_category`:

- `pre_change_category`: what was removed or lost. Options:
  `deforestation`, `urban_erosion`, `wetland_loss`, `water_contract`,
  `removed_crop_structure`.
- `post_change_category`: what appeared. Options:
  `vegetation_growth`, `new_building`, `new_road`, `new_infrastructure`,
  `new_crop_field`, `new_aquafarm`, `site_clearing`, `water_expand`, `mining`,
  `new_crop_structure`, `selective_logging`, `landslide`, `settlement`.
- `same_change_category`: an event where the land cover category is disturbed
  but not permanently converted to a different category. Options:
  `agricultural_activity`, `wildfire`, `ice_motion`, `flooding`.

## Data collection

The annotations were collected in several phases, all verified manually in an
annotation UI showing monthly Sentinel-2 mosaics.

- Phase 1: we trained a model to predict per-year land cover, then applied it
  on many randomly sampled locationsover a ten-year period. Candidate change
  locations were proposed by comparing the per-year land cover predictions:
  specifically, we looked for places where the model was confident that the
  land cover change was one type for three years, then we ignored the next
  year in case the change is gradual, and then the model was confident that
  the land cover change was another type for three more years.
- Phase 2: output-based labeling (apply the model at scale, look for change
  predictions at a confidence threshold tuned for high-recall and low-precision,
  then label based on those).
- Phase 3: output-based labeling.
- Phase 4: similar to Phase 1 but use a per-pixel land cover model instead of
  one that has a larger spatial context. This way we ensure we can find very
  small-scale changes.
- Phase 5: output-based labeling.
- Phase 6: label random points in 0.1 x 0.1 degree tiles that are likely to have
  changes (e.g. agricultural activity, wildfire, flooding, new airport, etc.).
- Phase 7: output-based labeling.
- Phase 8: output-based labeling.
- Phase 9: output-based labeling and focus on rare transitions (src_land_cover,
  dst_land_cover) that didn't appear frequently in the labels collected so far.
- Phase 10: similar to Phase 4 but in subsaharan Africa, and oversampling the
  tree -> bare transitions.
- Phase 11: output-based labeling focused on mining predictions.
- Phase 12: output-based labeling focused on mining predictions.
- Phase 13: output-based labeling focused on mining predictions.
- Phase 14: output-based labeling focused on mining predictions.
- Phase 15: output-based labeling focused on mining predictions.
- Phase 16: label points based on a set of 10 datasets related to mining in different
  parts of sub-saharan Africa.
- Phase 17: output-based labeling.
- Phase 18: output-based labeling.

## Model output summary rasters

Alongside the annotations, we provide summary GeoTIFFs produced by running the
trained LCC model at scale. The world is divided into 32768x32768-pixel UTM
tiles at 10 m/pixel (327.68 km per side), and each processed tile yields one
file named `{EPSG code}_{col}_{row}_summary.tif`, where `col` and `row` are the
tile's top-left corner in the 10 m/pixel UTM pixel grid (multiply by the
resolution to get projection coordinates).

Each file is a 9-band uint8 GeoTIFF summarizing the model's per-pixel
predictions for one reference timestamp. Pixels the model did not predict
(e.g. no cloud-free imagery) are 0 in every band.

| Band | Name | Description |
| --- | --- | --- |
| 1 | `binary_change` | Probability of change, scaled to 0-255. |
| 2 | `pre_class` | Argmax class of the pre/same change-category head (see table below). |
| 3 | `post_class` | Argmax class of the post change-category head (see table below). |
| 4 | `src_class` | Argmax source land cover class (see table below). |
| 5 | `dst_class` | Argmax destination land cover class (see table below). |
| 6 | `pre_score` | Probability (0-255) of the argmax class in band 2. |
| 7 | `post_score` | Probability (0-255) of the argmax class in band 3. |
| 8 | `ts_pre_month` | Predicted change start (last pre-change date), month-encoded. |
| 9 | `ts_post_month` | Predicted change end (change complete date), month-encoded. |

**Month encoding** (bands 8-9): 0 means no prediction; otherwise the value is
1 + the number of whole calendar months between January 2015 and the predicted
date. So 1 = January 2015, 13 = January 2016, and so on (255 reaches March
2036).

**Land cover classes** (bands 4-5): 0 = no prediction, then
1 = `bare`, 2 = `burnt`, 3 = `crops`, 4 = `fallow/shifting cultivation`,
5 = `grassland`, 6 = `Lichen and moss`, 7 = `shrub`, 8 = `snow and ice`,
9 = `tree`, 10 = `urban/built-up`, 11 = `water`, 12 = `wetland (herbaceous)`.

**Pre/same change-category classes** (band 2): in the deployed model, the
`pre_change_category` and `same_change_category` heads are merged into a
single head, with the "same" categories appended after the "pre" categories:

| Value | Category |
| --- | --- |
| 0 | no prediction |
| 1 | none |
| 2 | deforestation |
| 3 | urban_erosion |
| 4 | wetland_loss |
| 5 | water_contract |
| 6 | removed_crop_structure |
| 7 | agricultural_activity |
| 8 | wildfire |
| 9 | ice_motion |
| 10 | flooding |

**Post change-category classes** (band 3):

| Value | Category |
| --- | --- |
| 0 | no prediction |
| 1 | none |
| 2 | vegetation_growth |
| 3 | new_building |
| 4 | new_road |
| 5 | new_infrastructure |
| 6 | new_crop_field |
| 7 | new_aquafarm |
| 8 | site_clearing |
| 9 | water_expand |
| 10 | mining |
| 11 | new_crop_structure |
| 12 | selective_logging |
| 13 | landslide |
| 14 | settlement |

Note that "none" (value 1) is a valid prediction for both category heads: it
indicates the model predicts a change at that pixel but none of the listed
categories apply (or, for the merged pre/same head, that the change is fully
described by the post category, and vice versa). The argmax in bands 2-5 is
computed excluding the internal `nodata` class, so value 0 always means "no
prediction" rather than a predicted class.

For most applications, threshold band 1 (`binary_change`) first (e.g. at 128,
corresponding to probability 0.5) and interpret the remaining bands only at
pixels that pass the threshold; bands 2-9 are populated at every predicted
pixel regardless of whether change is predicted there.

For applications interested in specific types of changes that correspond to a
change category, filter for the class and threshold the corresponding
probability. For example, to create a mining mask, filter for band 3 `post_class = 10`
(10 is mining) and then threshold the band 7 probabilty.

## Related resources

- Training and inference code: [WIP](https://github.com/allenai/rslearn_projects/pull/294).
- Encoder: [OlmoEarth-v1.2-Base](https://huggingface.co/allenai/OlmoEarth-v1_2-Base).
- Imagery: Sentinel-2 L2A (Copernicus).