💧 Water Segmentation — Hard Boundary Detection

Kaggle Competition: Find the Water Task: Pixel-level segmentation of water regions with precise boundary detection Training time: ~1 hour on Kaggle T4 GPU

---

🏆 Final Results

Metric	Score
Val Dice	0.9814
Val IoU	0.9641
Composite Score	0.9710
Best Epoch	45 / 60
Optimal Threshold	0.60

---

📈 Training Curves

The model converged rapidly — hitting 0.94+ composite by epoch 5 — then continued refining steadily through epoch 45 before plateauing. No overfitting was observed: val loss tracked train loss closely throughout all 60 epochs.

Key milestones:

Epoch	Val Dice	Val IoU	Composite
1	0.858	0.763	0.801
5	0.963	0.931	0.944
10	0.971	0.946	0.956
20	0.978	0.958	0.966
43	0.981	0.964	0.971 ← best
60	0.978	0.959	0.967

---

🔍 Sample Predictions

Each row shows: input image → ground truth mask → model prediction. The model cleanly delineates water boundaries even in complex scenes with reflections, partial occlusion, and varying lighting.

---

📉 Threshold Optimization

Rather than using a fixed threshold of 0.5, the optimal binarization threshold was searched across the validation set using the actual competition metric. Higher thresholds produced sharper, more conservative boundaries — improving Boundary IoU and Contour F1 at the cost of slightly lower Region IoU.

Threshold	Score	B-IoU	C-F1	R-IoU
0.30	0.421	0.145	0.429	0.955
0.40	0.453	0.186	0.468	0.959
0.50	0.478	0.221	0.494	0.962
0.60	0.493	0.243	0.508	0.962

Optimal threshold: 0.60 — the scoring heavily weights boundary precision (80% of score), so a higher threshold that tightens the predicted boundary outperforms the default 0.5.

---

🧠 Approach

Why this problem is hard

Standard segmentation metrics (IoU, Dice) reward getting the bulk of the region right and are forgiving of sloppy edges. This competition flips that — 80% of the score comes from boundary precision:

Component	Weight	What it measures
Boundary IoU	40%	Overlap of dilated boundary pixels
Contour F1	40%	Point-to-point contour matching within 3px
Region IoU	20%	Standard mask overlap

A model that perfectly segments the region but has rough edges will score poorly. Every design decision here was made with boundary sharpness in mind.

---

Model: U-Net + EfficientNet-B3

Input (512×640) → EfficientNet-B3 Encoder → U-Net Decoder → Binary Mask

U-Net was chosen for its skip connections, which pass full-resolution spatial features directly from encoder to decoder — preserving the fine edge detail that deep encoders compress away. This is essential when boundary precision accounts for 80% of the score.

EfficientNet-B3 as the encoder provides a strong pretrained feature extractor with a good balance of accuracy and compute efficiency.

Input resolution 512×640 (upscaled from original) gives the model more pixels to work with at boundaries, directly improving contour precision.

---

Loss Function: EnhancedBoundaryLoss

Standard BCE or Dice loss only cares about pixel-level region overlap — they give no special weight to boundaries. The custom loss combines four terms:

Component	Weight	Role
BCE Loss	20%	Baseline per-pixel supervision
Dice Loss	30%	Region-level overlap
Gradient Loss	30%	Penalizes differences in spatial gradients — directly targets boundary sharpness
Sobel Edge Loss	20%	MSE between predicted and ground truth edge maps via Sobel filter

The gradient and Sobel terms together account for 50% of the total loss, meaning the model is explicitly trained to match boundaries rather than just regions.

---

Training Setup

Parameter	Value	Reason
Optimizer	AdamW	Weight decay regularization helps generalization
Learning rate	1e-4	Conservative LR for stable boundary learning
Scheduler	CosineAnnealingWarmRestarts	Periodic LR resets help escape local minima
Batch size	2 (effective 12)	Gradient accumulation × 6 steps
Mixed precision	AMP	Faster training, lower memory
Early stopping	Patience 20	Stopped at epoch 60 (patience 15/20)

---

Augmentation Strategy

Augmentations were kept deliberately conservative — aggressive transforms distort boundaries and degrade the contour metric.

• Geometric: horizontal/vertical flips, mild rotation (±15°), small scale/shift
• Color: brightness/contrast, hue-saturation, gamma
• Noise: Gaussian noise/blur (light)
• Weather: rain, fog, shadow (20% probability — simulates real-world water scenes)
• Dropout: CoarseDropout for occlusion robustness

---

Inference: Test-Time Augmentation (TTA)

At inference, each image is predicted 4 times and the results averaged:

1. Original
2. Horizontal flip → flip back
3. Vertical flip → flip back
4. Both flips → flip back

Averaging reduces prediction variance at boundaries, which directly improves Contour F1.

---

Post-Processing

1. Noise removal — connected components smaller than 100px are discarded as false positives
2. Morphological closing — small holes in the mask are filled with a 2×2 elliptical kernel
3. Edge refinement — in the boundary region (dilated − eroded mask), a slightly lower threshold (0.50 instead of 0.60) is applied to recover high-confidence edge pixels that the main threshold would discard

---

📁 Repository Structure

├── notebook.ipynb # Full training + inference notebook
├── submission.csv # Final predictions (RLE encoded, 200 images)
├── results/
│ ├── training_curves.png # Loss & Dice/IoU over 60 epochs
│ ├── predictions_sample.png # Image | Ground truth | Prediction
│ └── threshold_analysis.png # Competition score vs threshold
├── data/
│ └── threshold_optimization.csv # Raw threshold search results
└── README.md

---

🔁 Reproducing

1. Add the find-the-water dataset to your Kaggle notebook
2. Run all cells in notebook.ipynb
3. Outputs saved to /kaggle/working/: submission.csv, checkpoints/best_model.pth, training_history.csv

Training takes approximately 1 hour on a Kaggle T4 GPU (60 epochs × ~63s/epoch).