Kaggle Competition - DSTL Satellite Imagery Feature Detection
Tools:
Python + Keras with TensorFlow backend
OpenCV / Rasterio / Shapely for polygon manipulation.
Ideas:
-
XGBoost: pixel-based model
-
Unsupervised learning for waterway class: Canopy chloropyll content index (CCCI)
-
U-net for image segmentation:
-
Assembling with different input patch size
-
Cropping the output size guided by prediction loss
-
Incroporating evaluation metric into the loss function
-
Approximating Jaccard index such that it is differentiable
-
Mirror padding for global boundaries
-
Adding augmentation with rotation and reflection
-
Oversampling on rare classes (sliding steps)
-
Post-processing on standing water versus waterways, and small versus large vehicles.