Datasets:

timm
/

amazon-from-space

labels listlengths 1 9	label_names listlengths 1 9	image_id stringlengths 7 11	fold int8 1 9
[ 10, 12 ]	[ "haze", "primary" ]	train_0	5
[ 0, 5, 12, 16 ]	[ "agriculture", "clear", "primary", "water" ]	train_1	1
[ 5, 12 ]	[ "clear", "primary" ]	train_2	2
[ 5, 12 ]	[ "clear", "primary" ]	train_3	4
[ 0, 5, 9, 12, 13 ]	[ "agriculture", "clear", "habitation", "primary", "road" ]	train_4	7
[ 10, 12, 16 ]	[ "haze", "primary", "water" ]	train_5	8
[ 0, 5, 8, 12, 16 ]	[ "agriculture", "clear", "cultivation", "primary", "water" ]	train_6	3
[ 10, 12 ]	[ "haze", "primary" ]	train_7	7
[ 0, 5, 8, 12 ]	[ "agriculture", "clear", "cultivation", "primary" ]	train_8	8
[ 0, 5, 12, 15, 16 ]	[ "agriculture", "clear", "primary", "slash_burn", "water" ]	train_10	9
[ 5, 12, 16 ]	[ "clear", "primary", "water" ]	train_11	7
[ 6 ]	[ "cloudy" ]	train_12	6
[ 6 ]	[ "cloudy" ]	train_14	8
[ 5, 12 ]	[ "clear", "primary" ]	train_15	2
[ 5, 12 ]	[ "clear", "primary" ]	train_16	1
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_17	2
[ 5, 12 ]	[ "clear", "primary" ]	train_18	3
[ 0, 5, 12, 13 ]	[ "agriculture", "clear", "primary", "road" ]	train_19	1
[ 0, 5, 12, 16 ]	[ "agriculture", "clear", "primary", "water" ]	train_20	4
[ 5, 12, 13, 16 ]	[ "clear", "primary", "road", "water" ]	train_21	3
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_22	9
[ 0, 5, 12, 13 ]	[ "agriculture", "clear", "primary", "road" ]	train_23	2
[ 7, 11, 12 ]	[ "conventional_mine", "partly_cloudy", "primary" ]	train_24	3
[ 5, 12 ]	[ "clear", "primary" ]	train_25	3
[ 6 ]	[ "cloudy" ]	train_26	9
[ 0, 11, 12, 16 ]	[ "agriculture", "partly_cloudy", "primary", "water" ]	train_27	4
[ 5, 12 ]	[ "clear", "primary" ]	train_28	6
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_29	9
[ 5, 8, 12 ]	[ "clear", "cultivation", "primary" ]	train_31	3
[ 0, 5, 12, 13 ]	[ "agriculture", "clear", "primary", "road" ]	train_32	3
[ 0, 8, 11, 12 ]	[ "agriculture", "cultivation", "partly_cloudy", "primary" ]	train_33	5
[ 5, 12 ]	[ "clear", "primary" ]	train_34	7
[ 5, 12 ]	[ "clear", "primary" ]	train_35	7
[ 0, 5, 12, 13 ]	[ "agriculture", "clear", "primary", "road" ]	train_36	7
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_37	7
[ 5, 12 ]	[ "clear", "primary" ]	train_38	5
[ 5, 12 ]	[ "clear", "primary" ]	train_39	1
[ 0, 5, 8, 12 ]	[ "agriculture", "clear", "cultivation", "primary" ]	train_40	2
[ 0, 5, 12 ]	[ "agriculture", "clear", "primary" ]	train_41	3
[ 5, 12 ]	[ "clear", "primary" ]	train_42	4
[ 2, 10 ]	[ "bare_ground", "haze" ]	train_43	9
[ 5, 12 ]	[ "clear", "primary" ]	train_44	6
[ 5, 12 ]	[ "clear", "primary" ]	train_45	2
[ 2, 5, 9, 12 ]	[ "bare_ground", "clear", "habitation", "primary" ]	train_46	9
[ 5, 12, 16 ]	[ "clear", "primary", "water" ]	train_48	8
[ 5, 12 ]	[ "clear", "primary" ]	train_49	8
[ 6 ]	[ "cloudy" ]	train_50	2
[ 5, 12 ]	[ "clear", "primary" ]	train_51	3
[ 0, 5, 9, 12, 13, 16 ]	[ "agriculture", "clear", "habitation", "primary", "road", "water" ]	train_52	2
[ 5, 12 ]	[ "clear", "primary" ]	train_53	5
[ 5, 12 ]	[ "clear", "primary" ]	train_54	8
[ 11, 12, 13, 16 ]	[ "partly_cloudy", "primary", "road", "water" ]	train_55	7
[ 5, 12, 16 ]	[ "clear", "primary", "water" ]	train_56	4
[ 1, 2, 5, 12, 16 ]	[ "artisinal_mine", "bare_ground", "clear", "primary", "water" ]	train_57	3
[ 0, 5, 12, 13 ]	[ "agriculture", "clear", "primary", "road" ]	train_58	5
[ 0, 5, 9, 12, 13 ]	[ "agriculture", "clear", "habitation", "primary", "road" ]	train_59	1
[ 10, 12 ]	[ "haze", "primary" ]	train_60	4
[ 5, 12 ]	[ "clear", "primary" ]	train_61	9
[ 5, 12, 16 ]	[ "clear", "primary", "water" ]	train_62	3
[ 0, 5, 12, 13 ]	[ "agriculture", "clear", "primary", "road" ]	train_63	5
[ 0, 5, 12 ]	[ "agriculture", "clear", "primary" ]	train_64	3
[ 5, 12 ]	[ "clear", "primary" ]	train_65	1
[ 0, 5, 12, 13 ]	[ "agriculture", "clear", "primary", "road" ]	train_66	5
[ 5, 9, 12, 13 ]	[ "clear", "habitation", "primary", "road" ]	train_67	8
[ 6 ]	[ "cloudy" ]	train_68	1
[ 5, 12 ]	[ "clear", "primary" ]	train_69	1
[ 3, 5, 12 ]	[ "blooming", "clear", "primary" ]	train_70	5
[ 5, 12, 16 ]	[ "clear", "primary", "water" ]	train_71	5
[ 5, 12 ]	[ "clear", "primary" ]	train_73	6
[ 0, 11, 12, 13 ]	[ "agriculture", "partly_cloudy", "primary", "road" ]	train_74	4
[ 0, 5, 12, 13, 16 ]	[ "agriculture", "clear", "primary", "road", "water" ]	train_75	5
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_76	2
[ 6 ]	[ "cloudy" ]	train_77	3
[ 5, 8, 12 ]	[ "clear", "cultivation", "primary" ]	train_79	6
[ 0, 11, 12, 13 ]	[ "agriculture", "partly_cloudy", "primary", "road" ]	train_80	4
[ 5, 12 ]	[ "clear", "primary" ]	train_81	9
[ 5, 12 ]	[ "clear", "primary" ]	train_83	5
[ 0, 5, 9, 12, 13 ]	[ "agriculture", "clear", "habitation", "primary", "road" ]	train_84	1
[ 5, 12, 16 ]	[ "clear", "primary", "water" ]	train_85	4
[ 6 ]	[ "cloudy" ]	train_87	4
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_88	8
[ 5, 12 ]	[ "clear", "primary" ]	train_89	2
[ 0, 11, 12, 13 ]	[ "agriculture", "partly_cloudy", "primary", "road" ]	train_90	4
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_91	4
[ 0, 5, 9, 12, 13 ]	[ "agriculture", "clear", "habitation", "primary", "road" ]	train_92	1
[ 5, 12 ]	[ "clear", "primary" ]	train_93	3
[ 0, 5, 9, 12, 13 ]	[ "agriculture", "clear", "habitation", "primary", "road" ]	train_94	3
[ 5, 12 ]	[ "clear", "primary" ]	train_95	6
[ 5, 12 ]	[ "clear", "primary" ]	train_96	8
[ 6 ]	[ "cloudy" ]	train_97	1
[ 5, 12 ]	[ "clear", "primary" ]	train_98	3
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_99	2
[ 2, 5, 16 ]	[ "bare_ground", "clear", "water" ]	train_100	9
[ 6 ]	[ "cloudy" ]	train_101	3
[ 5, 12 ]	[ "clear", "primary" ]	train_102	1
[ 0, 5, 9, 12, 13, 16 ]	[ "agriculture", "clear", "habitation", "primary", "road", "water" ]	train_103	7
[ 10, 12 ]	[ "haze", "primary" ]	train_104	2
[ 5, 12 ]	[ "clear", "primary" ]	train_105	8
[ 5, 12 ]	[ "clear", "primary" ]	train_106	2
[ 11, 12 ]	[ "partly_cloudy", "primary" ]	train_107	5

End of preview. Expand in Data Studio

Description

Dataset from the Kaggle Planet: Understanding the Amazon from Space competition (2017). This is the JPG (visual RGB) variant of the release.

The images are 256x256 pixel chips cut from Planet's PlanetScope scenes of the Amazon basin (~3 m ground sample distance). Each chip is tagged with one atmospheric label and zero or more land cover / land use labels.

Labels (17):

atmospheric (exactly one per chip): clear, partly_cloudy, cloudy, haze. Chips tagged cloudy have no other labels.
common land: primary (primary rainforest), water, agriculture, road, habitation, cultivation (shifting / small-scale cultivation), bare_ground.
rare land: slash_burn, selective_logging, blooming, conventional_mine, artisinal_mine (sic, original spelling kept), blow_down.

The original dataset has one labelled train split (train_v2.csv) and an unlabelled test set whose labels were hidden for the competition and have not been released. The test split here contains all 61,191 test JPGs (40,669 test_* + 20,522 file_* from the "test-additional" set), with labels, label_names and fold set to null.

I have split the 40,479 labelled train images into 10 folds using multi-label iterative stratification (Sechidis et al., 2011, seed 42). Every label, including the rare ones, is spread near-evenly across the folds.

Website: https://www.kaggle.com/c/planet-understanding-the-amazon-from-space

Configs

default - train (folds 1-9, 36,431 images), validation (fold 0, 4,048 images), test (61,191 unlabelled images).
folds - train (all 10 folds, 40,479 images) and test. Filter on the fold column for k-fold cross validation. This config reads the same parquet files as default, so the images are not duplicated.

from datasets import load_dataset
ds = load_dataset('timm/amazon-from-space', 'folds', split='train')
train_ds = ds.filter(lambda f: f != 3, input_columns='fold')
val_ds = ds.filter(lambda f: f == 3, input_columns='fold')

Dataset Structure

image - JPEG image, 256x256 pixels, the original bytes from the source release. Note: the source JPGs are stored as 4-channel CMYK (inverted RGB with an empty K channel). PIL decodes them to CMYK mode, and image.convert('RGB') gives the correct colours. timm and most PIL-based pipelines already do this conversion. With other decoders, check the colours.
labels - sequence of ClassLabel indices of the positive labels.
label_names - names of the positive labels.
image_id - original image filename stem (e.g. train_1234, test_42, file_7).
fold - stratified fold index (0-9) for labelled images, null for test.

Usage

This dataset follows the same multi-label format as timm/plant-pathology-2021 for use with timm 1.0.31 or greater:

python train.py --dataset hfds/timm/amazon-from-space --train-split train --val-split validation \
  --task multilabel --target-key labels --num-classes 17 \
  --model resnet50.ram_in1k --pretrained --batch-size 64 --epochs 30 \
  --opt adamw --lr 1e-4 --weight-decay 0.01 --amp

python validate.py --dataset hfds/timm/amazon-from-space --split validation \
  --task multilabel --target-key labels --num-classes 17 \
  --model resnet50 --checkpoint ./output/train/<experiment>/model_best.pth.tar

The competition metric was the sample-averaged F2 score.

Citation

Planet: Understanding the Amazon from Space.
https://kaggle.com/competitions/planet-understanding-the-amazon-from-space, 2017. Kaggle.

Downloads last month: -