Task: semantic segmentation of satellite RGB imagery with only 72 annotated samples of varying dimensions. I first reconstructed 8 original satellite tiles and generated a larger training set by adaptive/dynamic tiling into ~12k patches of size 128x128, minimizing bilinear downsampling artifacts and reducing compute.

To improve sample efficiency vs. a pure U-Net approach, I built a hybrid pipeline: image enhancement using Non-Local Bayes denoising + SOS (Second-Order Statistics) refinement, short-term wavelet denoising, Richardson-Lucy deconvolution, and CLAHE (Contrast Limited Adaptive Histogram Equalization), followed by Geodesic Voronoi Diagram superpixel segmentation (color-aware geodesic distance) inspired by Zeng et al. (ICCV 2011). Centroids were initialized using 16-color quantization plus flood-fill region growing to build an intra-region grid and obtain partially color-aware centroid placement, avoiding slow iterative centroid optimization.

I extracted per-superpixel features (min/max/mean/std/skewness over R,G,B) and trained an XGBoost classifier to label superpixels. Predictions were merged back to tile scale by fusing superpixels, then post-processed with label smoothing / spatial consistency filtering to reduce isolated errors and realigned from superpixels to the original tiles.