Chesapeake Land Cover - Subset

A small subset of the Microsoft Chesapeake Land Cover dataset that is ML-ready for semantic segmentation.

Original data: https://lila.science/datasets/chesapeakelandcover

Responsible party for this data subset: Andrew H. Fagg

Dataset Contents

Both the tar and zip files contain the same material: three different TF-Dataset directories:

• tf_dataset_train_F0
• tf_dataset_valid_F0
• tf_dataset_test_F0

Each TF-Dataset is a sequence of individual examples composed of 2-tuples:

1. 256 x 256 x 26: float32 features for each 1m x 1m pixel (values are in the range 0 .. 1).
2. 256 x 256: int8 semantic labels for each pixel.

The features are a composition of data from several different satellites, including two visible light RGB images, and a range of infrared images.

Each pixel has a semantic label:

• 0 = no class (unknown)
• 1 = water
• 2 = tree canopy / forest
• 3 = low vegetation / field
• 4 = barren land
• 5 = impervious (other)
• 6 = impervious (road)

Data Origins

The examples are all derived from Fold 0 of the original dataset:

• The subsetted training set is all "train" examples whose ids end in 0: *0.npz (500 examples)
• The subsetted validation set is all "train" examples whose ids end in 9: *9.npz (500 examples)
• The subsetted testing set is all "valid" examples: *.npz (250 examples)

Dataset Use

This subsetted dataset can be used with any model that makes use of TF-Dataset formatted data for training and/or evaluation. In particular, this dataset is used with the semantic segmentation example from Zero2Neuro