Multiclass segmentation of cloud patterns in satellite images.
![]() |
![]() |
---|
Code for Kaggle's Understanding Clouds from Satellite Images Challenge. This solution has scored 0.65455 (Dice) on private leaderboard and ranked 161st place (top 11%).
-
PyTorch
-
Ensemble of 4 pretrained models from Segmentation Models (Unet, FPN) and FastFCN (EncNet, DeepLabV3)
-
Optimizer: RAdam (lr=0.005)
-
Metric: Dice
-
Loss: Focal + log dice
-
Scheduler: ReduceLROnPlateau
-
Preprocessing: Resize from 1400x2100 to 320x480
-
Postprocessing: Threshold, Remove small masks, Draw convex hull mask
-
Albumentations augmentations
-
Trained on 80% of data (stratified based on mask areas)
-
35 epochs, models with best Dice score were selected for ensemble
Local dice: 0.64465
Public leaderboard score: 0.65574
Private leaderboard score: 0.65455
Model | Backbone | Local Dice |
---|---|---|
FPN | efficientnet-b4 | 0.46 |
Unet | efficientnet-b4 | 0.444 |
EncNet | resnet50 | 0.445 |
DeepLabV3 | resnet50 | 0.436 |
Class | Fish | Flower | Gravel | Sugar |
---|---|---|---|---|
Local Dice | 0.578186 | 0.750014 | 0.642398 | 0.608013 |