Spaces:
Sleeping
Sleeping
Commit History
-
Redesign: five-step pedagogical flow with spectral baseline
functionNormallyStep 1 - Know Your Data: full-scene composite viewer (natural/false/SWIR/single-band) with labelled-pixel markers overlaid, clickable pixel inspection. Step 2 - Spectral Signatures: per-class mean±sigma chart and NDVI/NDWI heatmaps computed from raw (unnormalised) bands for correct ratio semantics. Step 3 - Spectral Baseline: pure-numpy chunked KNN classifier that predicts the full 501x1001 scene from 7-band spectral features alone (no spatial context). Step 4 - Deep Learning: UNet training with side-by-side ground-truth / KNN / UNet patch comparison to make the spatial-context benefit concrete. Step 5 - Experiment Lab: compare up to 3 UNet experiments; max capped at 3. Also fixes natural-colour composite to use H4/H3/H2 (Red/Green/Blue) instead of the previous H3/H2/H1, and adds matplotlib + scipy to requirements.089078d
-
Fix mask interpretation: treat value 0 as background, not Water
functionNormallyJulien's data uses 1-indexed classes (1=Water 2=Urban 3=Agriculture 4=Forest). Value 0 is unlabeled background and must map to IGNORE_INDEX. Previous code treated 0 as Water, causing 99.9% of pixels to appear as Water class. Also simplify patch extraction to include any patch with >=1 labeled pixel, removing the 15% threshold that was impossible to satisfy with only ~500 sparse point annotations in the entire image.809b793 -
Fix spatial size mismatch for patch sizes larger than image dimensions
functionNormally- Model: interpolate output to input spatial size when U-Net pooling shrinks non-divisible-by-8 inputs - Data: pad last-resort fallback patch to full patch_size instead of returning a smaller crop that would mismatch target15f6af6 -
Add per-class pixel distribution to diagnostics, raise patch slider to 512
functionNormally3ed52e7 -
Fix patch extraction to always sample image edges
functionNormallyWith H=501 and stride=128, range(0,246,128)=[0,128] skipped y=245, so rows 245-500 were never included. Non-water classes living in those rows caused everything to appear as Water.3d56599 -
Robust mask remapping: auto-detect class values, add diagnostics, fix shape mismatch
functionNormally2c61116 -
Fix mask filenames: TRAINING.tif / GROUND TRUTH.tif
functionNormallye32a8fd -
Switch to real dataset: 4-class, rasterio patch loader, ignore unlabeled pixels
functionNormally9cb9bec
-
Upload 7 files
db426cb -
Delete data.py
30a7194 -
Upload 6 files
3ab9bb3