clarkcga / multi-temporal-crop-classification-training-data Goto Github PK

Using the definition of tiles from here, we need to retrieve corresponding scenes from HLS dataset.
Specification:

• Find three best cloud-free scenes during the growing season. (growing season needs to be defined after we decide on what crop types to include, 1)
• export the scenes for each chip. Scenes should include all the bands from HLS data at 30m spatial resolution.

Both HLS and CDL are at 30m spatial resolution.
We need to confirm if they are using the same coordinate system and grid. If not, decide how to project one to the other. Preferably, we should project HLS to CDL as CDL is a categorical data layer and more noise will be added to the dataset if we reproject it.

The outcome of this task should be:

• the final AOI for our training dataset, which can be a subset of the GFM AOI. (the AOI needs to be a multiply of 224 x 224 pixel chips)
• The projection and coordinate system for the training dataset

Retrieve CDL data (latest year available) for the region that GFM is trained on (AOI).
If needed, crop the dataset to the AOI and store under /data.

Using the tiles defined here, chip the CDL dataset and export a GeoTIFF file corresponding to the HLS tile.

Divide the training dataset AOI (from here) to tiles of 224 x 224 pixels (at 30m spatial resolution).
The outcome can be a GeoJSON which stores all of these tiles, or any other compatible data format.
We will use the bounding boxes of these tiles to chip the CDL dataset and query HLS scenes.