| # README |
|
|
| ## Overview |
|
|
| This dataset was designed to provide a framework and set of evaluations for |
| assessing the performance of geospatial models and datasets on real-world, |
| data-limited applications. It is comprised of fifteen individual evaluations |
| derived from ten publicly-available data sources. These datasets were selected |
| and processed to represent archetypal classification, regression, and change |
| detection use cases. A list of short names used to refer to individual |
| evaluation datasets, names of the sources these datasets were derived from, and |
| path to the location within this archive are provided in the table below. The |
| following sections provide a basic overview of general pre-processing, and a |
| README file is provided for each individual data source and dataset. |
|
|
| Short name | Source name | Path |
| --------------------- | ----------------------------------------------------------------------------------------------- | ---- |
| `africa_crop_mask` | Comparison of Cropland Maps Derived from Land Cover Maps in Sub-Saharan Africa | `./africa_crop_mask` |
| `aster_ged` | AG100: ASTER Global Emissivity Dataset 100-meter V003 | `./aster_ged` |
| `canada_crops_coarse` | Agriculture and Agri-Food Canada (AAFC) Annual Crop Inventory Ground Truth Data | `./canada_crops/canada_crops_coarse` |
| `canada_crops_fine` | Agriculture and Agri-Food Canada (AAFC) Annual Crop Inventory Ground Truth Data | `./canada_crops/canada_crops_fine` |
| `descals` | Global oil palm extent and planting year from 1990 to 2021 | `./descals` |
| `ethiopia_crops` | Ethiopian Crop Type 2020 (EthCT2020) dataset | `./ethiopia_crops` |
| `glance` | GLanCE: A Global Land Cover Training Dataset from 1984 to 2020 | `./glance` |
| `lcmap_lc` | LCMAP CONUS Reference Data Product 1984-2021 land cover, land use and change process attributes | `./lcmap/lcmap_lc` |
| `lcmap_lu` | LCMAP CONUS Reference Data Product 1984-2021 land cover, land use and change process attributes | `./lcmap/lcmap_lu` |
| `lcmap_lcc` | LCMAP CONUS Reference Data Product 1984-2021 land cover, land use and change process attributes | `./lcmap/lcmap_lcc` |
| `lcmap_luc` | LCMAP CONUS Reference Data Product 1984-2021 land cover, land use and change process attributes | `./lcmap/lcmap_luc` |
| `lucas_lc` | LUCAS: Land Use and Coverage Area frame Survey | `./lucas/lucas_lc` |
| `lucas_lu` | LUCAS: Land Use and Coverage Area frame Survey | `./lucas/lucas_lu` |
| `openet_ensemble` | OpenET Ensemble Monthly Evapotranspiration v2.0 | `./openet_ensemble` |
| `us_trees` | Research-grade iNaturalist observation filtered to tree genera in the United States | `./us_trees` |
|
|
| ## Dataset format |
|
|
| All evaluation datasets were processed to the standard format and properties |
| described in the table below. Key properties include location coordinates, |
| label, and time period over which that label is considered applicable or |
| "valid". In some cases, e.g., LCMAP, LUCAS, and Canada crops, multiple |
| evaluation datasets were created using different hierarchies or combinations of |
| labels. For regression datasets, there is no `label_name` column, and the |
| `label` column records the measurement value. Change detection datasets include |
| a number of additional columns to characterize "before" and "after" periods and |
| labels, in addition to a general "change" / "no change" label. |
|
|
| Table: Evaluation dataset properties. |
|
|
| | Property | Units | Notes | |
| | :----------------------------------- | :-------- | :------------------------ | |
| | x | decimal | Longitude coordinate. | |
| : : degrees : Must have at least 10\-4 : |
| : : : precision to be : |
| : : : considered valid at \~10m : |
| : : : resolution. : |
| | y | decimal | Latitude coordinate. Must | |
| : : degrees : have at least 10\-4 : |
| : : : precision to be : |
| : : : considered valid at \~10m : |
| : : : resolution. : |
| | label | numeric | Column recording the | |
| : : (int or : label or measurement : |
| : : float) : field used for evaluation : |
| : : : (one per dataset). For : |
| : : : classification, this : |
| : : : should be dense : |
| : : : sequential remapping of : |
| : : : ‘label\_name’ values. For : |
| : : : regression, this should : |
| : : : be the (continuous) : |
| : : : measured/observed value : |
| | label\_name | str | (optional) This field can | |
| : : : be used to preserve : |
| : : : values/codes from the : |
| : : : original dataset for : |
| : : : human-readability / : |
| : : : easier cross-checking and : |
| : : : visualization. Strongly : |
| : : : recommended for : |
| : : : classification evals, not : |
| : : : required for regression : |
| : : : evals. : |
| | valid\_time\_start\_ms | millis | Start and end times | |
| : valid\_time\_end\_ms : : defining the range over : |
| : : : which the : |
| : : : label/measurement is : |
| : : : valid (may be same for : |
| : : : instantaneous/single-date : |
| : : : measurements). This is : |
| : : : the period over-which the : |
| : : : embedding summary is : |
| : : : created and should not : |
| : : : extend more than 6 months : |
| : : : before or after the : |
| : : : support period. : |
| | support\_time\_start\_ms | millis | Start and end times | |
| : support\_time\_end\_ms : : defining a support period : |
| : : : for informing prediction. : |
| : : : This is the period over : |
| : : : which input data is : |
| : : : fetched for each row. It : |
| : : : must be no longer than 1 : |
| : : : year in length. : |
| | split | str | Each label/observation | |
| : : (train or : (row) should be assigned : |
| : : test) : to a fixed train/test : |
| : : : split. : |
| | shard | numeric | {optional) Assign a shard | |
| : : (int) : to each row for efficient : |
| : : : ingestion. A shard should : |
| : : : be associated with no : |
| : : : more than 2000 rows. : |
| | *label\_before\** | numeric | Integer label for | |
| : : (int) : “before” class. : |
| | *label\_before\_name\** | str | This is used to preserve | |
| : : : “before” values/codes : |
| : : : from the original dataset : |
| : : : for human-readability / : |
| : : : easier cross-checking and : |
| : : : visualization : |
| | *label\_after\** | numeric | Integer label for “after” | |
| : : (int) : class : |
| | *label\_after\_name\** | str | This is used to preserve | |
| : : : “after” values/codes from : |
| : : : the original dataset. : |
| | *valid\_time\_start\_before\_ms\** | millis | Start and end times | |
| : *valid\_time\_end\_before\_ms\** : : defining the range over : |
| : *valid\_time\_start\_after\_ms\** : : which the “before” and : |
| : *valid\_time\_end\_after\_ms\** : : “after” : |
| : : : labels/measurements are : |
| : : : valid (may be same for : |
| : : : instantaneous/single-date : |
| : : : measurements). These are : |
| : : : the periods over-which : |
| : : : embedding summaries are : |
| : : : created. : |
| | *support\_time\_start\_before\_ms\** | millis | Start and end times | |
| : *support\_time\_end\_before\_ms\** : : defining the before and : |
| : *support\_time\_start\_after\_ms\** : : after change support : |
| : *support\_time\_end\_after\_ms\** : : periods. : |
|
|
| \* *These fields are required for change detection datasets only* |
|
|
| ## General pre-processing |
|
|
| In selecting evaluation datasets, we preferred point measurements or annotations |
| over polygons, given that reasoning about labels for a specific point rather |
| than over a larger area is more straightforward, and this simplifies sampling of |
| feature vectors and/or mapped results for comparisons across approaches. We only |
| selected datasets where point coordinate data (longitude, latitude) in decimal |
| degrees was sufficiently precise relative to a nominal 10-meter resolution, |
| i.e., at least four decimal points of precision (0.0001), which is about 11.1m |
| at the equator. Point observations within each dataset were filtered to |
| guarantee a minimum distance of 1.28 km between sampled points in order to |
| reduce spatial autocorrelation between training and test sites (and this process |
| will be hereafter referred to as “spatial proximity filtering”). Given our focus |
| on temporal precision, we also required that labels have a clearly defined |
| “valid period” over which the label could be reasonably applied, i.e., a range |
| or instant (annual, monthly, single-date), and this period must intersect 2017 |
| onward. We ensured that we had representation of different temporal aggregations |
| across our final set of evaluations. Sample points were allocated to train and |
| test splits such that the training datasets were balanced by class (or regularly |
| spaced bins in the case of regression datasets), with no per-class sample |
| exceeding 300 points and the remainder of the points allocated to an unbalanced |
| test split. When possible, we used existing Google Earth Engine assets for |
| publicly available datasets; otherwise, reference datasets were downloaded from |
| archived sources. Additional details on sources and processing for individual |
| datasets are provided in README files provided in nested directories. |
|
|
| ## Trial group configuration |
|
|
| Each dataset has an associated trial group configuration file. This is a |
| JSON-syntax file that specifies which groups of trials should be performed |
| during evaluation. For example, one configuration from |
| africa_crop_mask_trial_groups.json) is `json { "id": |
| "africa_crop_mask_knn3_500x10", "repeats_n": 500, "samples_n": 10, "model_type": |
| "knn", "metrics_type": "classification", "model_config": {"k": 3} }` |
| |
| This says that during the evaluation of the Sub-Saharan Africa Cropland Maps |
| dataset, a group of 500 classification trials should be run. Each trial in that |
| group will randomly select 10 samples from each class (selecting only from the |
| `train` split), train a KNN model with `k = 3`, and evaluate the performance of |
| that model on all the points in the `test` split. Evaluation of regression |
| datasets is similar, except that `samples_n` specifies how many samples should |
| be drawn from each histogram bin, using the `partition` column. |
|
|
| ## License and Disclaimer |
|
|
| Copyright 2025 Google LLC All software is licensed under the Apache License, |
| Version 2.0 (Apache 2.0); you may not use this file except in compliance with |
| the Apache 2.0 license. You may obtain a copy of the Apache 2.0 license at: |
| https://www.apache.org/licenses/LICENSE-2.0 |
|
|
| All other materials, except as set out below, are licensed under the Creative |
| Commons Attribution 4.0 International License (CC-BY). You may obtain a copy of |
| the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode. |
|
|
| * iNaturalist is licensed under a Creative Commons Attribution Non-Commercial |
| 4.0 License (CC-BY-NC). You may obtain a copy of the CC-BY-NC license at |
| https://creativecommons.org/licenses/by-nc/4.0/deed.en. |
| * Canadian AAFC Annual Crop Inventory from the Canadian AAFC (Agriculture and |
| Agri-Food Canada) is licensed under the Open Government Licence Canada. You |
| may obtain a copy of the licence at |
| https://open.canada.ca/en/open-government-licence-canada. |
| |
| Attribution for all datasets listed in the ‘Overview’ section above is set out |
| in their respective README files. |
|
|
| Unless required by applicable law or agreed to in writing, all software and |
| materials distributed here under the Apache 2.0 or CC-BY licenses are |
| distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, |
| either express or implied. See the licenses for the specific language governing |
| permissions and limitations under those licenses. |
|
|
| This is not an official Google product. |
|
|
