Model weights and training, validation, and test set images and masks for "Mapping one million small reservoirs in Brazil highlights widespread environmental and policy implications"

Annotated masks and Sentinel-1/-2 images split into training, validation, and test sets. Used for training convolutional neural network for small reservoir mapping. 

- manet_sentinel.ckpt: PyTorch model checkpoint file containing model weights.

- annotations_locations_stats.csv: Basic information about annotation masks, including names, locations, and pixel-wise accuracy statistics of final model. Contains columns:

1. name: File name, with .tif extension. Annotation file name is the same but with .png extension.
2. set: train/val/test
3. biome: Biome of center coordinates
4. center_longitude, center_latitude: Coordinates of center of annotation
5. xmin, xmax, ymin, ymax: Mask extents, in lat/lon coordinates
6. true_positive_pixels, false_positive_pixels, true_negative_pixels, false_negative_pixels: Pixel-wise statistics based on final model predictions.

- annotations.zip:  Contains binary reservoir masks (0 is non-reservoir, 1 is reservoir) split into training, validation, and test sets, in png format.

- images.zip: Contains Sentinel-1/-2 images in tif format split into training, validation, and test sets with the following bands:

1. Blue
2. Green
3. Red
4. Near-infrared
5. Sentinel-1 SAR VV
6. Sentinel-1 SAR VH
7. NDVI
8. NDWI
9. Gao's NDWI
10. MNDWI

- manually_checked_detections.gpkg: We also include a shapefile of 5000 manually checked reservoir detections including their accuracy assessment and error type. Accuracy assessment ("label") is one of "True Positive", "Questionable", and "False Positive". Error type ("error_type") is one of:

1. True Positive (i.e. not an error)
2. Natural Water Body
3. Non-Stream-Fed Impoundment (i.e. off-stream)
4. Duplicate Detection
5. Shadow
6. Land
7. Fragment of >50 ha reservoir