This document details the design for a machine learning training and evaluation pipeline to successfully fulfill a ground truth-based satellite imagery snow cover identification task.
Goal: produce a model which can identify snow in 3m PlanetScope satellite imagery with sufficient accuracy for operational use.
Our previous approach used a Gaussian Process approach to create a pixel-based snow classifier. The results were mixed. We're hoping to expand based on developments in image segmentation networks.
The framework we're using for this task is an "image segmentation" modeling framework, which relies on three elements:
- Images of known, consistent sizes
- Binary masks corresponding to the desired segmentation for each image, with matching sizes
- A network architecture (with or without pre-trained weights?) for image segmentation
| Data | Data Type | Description |
|---|---|---|
| PlanetScope Image Tiles | 4-band rasters | Using the ../preprocess tools we've created a set of 4-band TIFF image tiles which are stored in OSM/XYZ tile format and represent imagery relevant to the known ground-truth. |
| Ground Truth Tiles | Binary rasters | Again using ../preprocess toosl we've created a set of 1-band binary TIFF image tiles (in OSM/XYZ) which represent our segmentation masks. Note: These tiles cover an extent at least as large as the image tiles. The ground-truth data can come from a variety of sources, including: 1) the Airborne Snow Observatory (ASO) and 2) SnowEX |
TBD, but
We need to figure out how to structure this training task, because the data are strange. Though we have input data as above, it's really stored on disk in slippy-map tile directories for each Planet image. The options for incoporating these data into a training pipeline are as follows:
- Concatenate all tiles from all images into a
/image-tilesfolder.- Pros: training code / data loader is easy and straightforward.
- Cons: lose geospatial information in the form of XYZ directory structure, which makes serving results harder? Need to assign unique IDs to each image.
- Write a data loader which does this concatenation given a list of folders which themselves are slippy map directories.
- Pros: means we can keep the data stored in the current way (which is produced in the
preprocessstep) - Cons: would be pretty precise + brittle to implement this and to ensure overlap with the correct mask tile
- Pros: means we can keep the data stored in the current way (which is produced in the
- Train with checkpoints: train model incrementally using only a single image (e.g. a single XYZ directory) for each training step and using model checkpoints to continue training.
- Pros: allows for maintainance of the current directory structure and for easy train-test split (if we split on images!)
- Cons: there's got to be some sort of training bias lurking in here. Difficult train-test split (if we split on tiles!)
I've decided that the best first-pass at this is to just concatenate all images together into an /image-tiles folder, which has several implementation advantages. The loss of geoaptial information isn't terribly important (especially when the .tif files themselves are GeoTiffs.)
- Ecology: studying changes in alpine phenology (Janneke HilleRisLambers)
- Ecology: abiotic environmental variable for species distribution modeling
- Avalanche Forecasting: early season snow extent for basal weak layer prediction (Deems)