Original dataset. The original dataset contains satellite imagery from Landsat 7 as input data and collocated binary masks of kelp forest as target data. The input images contain 5 visual channels and 2 masks. Images and masks have a resolution of 350 x 350 pixels:
- Short-wave infrared (1550nm - 1750nm)
- Near infrared (770nm - 900nm)
- Red (630nm - 690nm)
- Green (520nm - 600nm)
- Blue (450nm - 520nm)
- Cloud cover mask
- Digital elevation model used as land/sea mask
Landsat 7 SLC issue. Landsat 7 has issues with its Scan Line Corrector (SLC)
since ~2005 (https://www.usgs.gov/faqs/what-are-landsat-7-slc-gap-mask-files).
That results in invalid values (NaN) in form of stripes on some of the images.
These invalid values are filled with OpenCV (see step_1a_impute.ipynb)
Remote sensing indices. The 5 original bands are used to compute the following
remote sensing (RS) indices as additional features (see step_1b_preprocess.ipynb):
- NDWI1
- NDWI2
- NDVI
- GNDVI
- NDTI
- EVI
- CARI
All indices are added as additional channels to the original imagery.
Normalization. After computation of the RS indices, each channel is normalized
to step_1b_preprocess.ipynb).
Two models are trained on versions of the same dataset to achieve different tasks. The motivation is to create two models of different complexities -- a simple one and a complex one --yielding better performance when combined compared to their individual performance. Additionally, the model outputs can be assessed separately by the user for agreement and sanity, which is viewed as operational/production-use benefit.
For robustness and uncertainty assessment, multiple similar models of both model types
are trained to form an ensemble of models. That allows us to predict a map of kelp
probabilities instead of just a binary mask. The ensemble is generated by training
step_3a_eval_ens.py: Individual inference for each member of
the ensemble for both model types.
step_3b_eval_ens_2staged.py: Combine individual
predictions into single prediction.
- Competition website: https://www.drivendata.org/competitions/255/kelp-forest-segmentation/