Satellite Image Segmentation using U-Net

Land-cover maps underpin urban planning, crop monitoring, deforestation tracking and disaster response, yet hand-labelling satellite imagery is slow, expensive and inconsistent between annotators — two people rarely draw the same boundary around a forest edge. Manual mapping cannot keep pace with how fast landscapes change, and it does not scale to large regions or repeat surveys. Semantic segmentation automates this pixel by pixel: a network assigns a land-cover class to every pixel of a satellite tile, producing a map a planner can review in minutes instead of weeks. The technical challenge is real — huge images that must be tiled, strong class imbalance between rare urban areas and vast barren land, and only a few hundred labelled images to learn from. That is exactly why the project is a strong academic build: it demonstrates transfer learning with a pre-trained encoder, class-imbalance handling with weighted losses, tiled inference with stitching, and honest segmentation metrics rather than pixel accuracy.

Dataset & model:
Dataset name: DeepGlobe Land Cover Classification.
Source: publicly released remote-sensing segmentation benchmark.
Task: semantic segmentation (per-pixel classification).
Classes: 7 — urban, agriculture, rangeland, forest, water, barren, unknown.
Model: U-Net with a pre-trained encoder (transfer learning); PyTorch/torchvision segmentation pipeline.
Input: satellite image tile (source images are 2448x2448 pixels at 50 cm per pixel), tiled into overlapping windows.
Prediction: per-pixel class probability map.
Output: land-cover mask overlaid on the input tile with a color legend; exportable as PNG.
Evaluation metrics: per-class IoU and mean Dice — computed by the training notebook on the validation split during the build.
Design target: ~70%+ mean IoU on the validation split. This is a design target, not a measured claim; the notebook computes the actual per-class IoU and Dice during the build.

Working:

1. Training phase: large 2448x2448 satellite images are split into overlapping tiles (e.g. 512x512), normalized, and augmented with rotations, flips and brightness jitter through a geospatial preprocessing pipeline.
2. A U-Net with a pre-trained encoder is fine-tuned on the DeepGlobe training split with weighted cross-entropy plus Dice loss, so under-represented classes like urban are not drowned out by barren and rangeland.
3. Evaluation phase: the training notebook evaluates the run on the validation split, computing per-class IoU and mean Dice and visualizing where the model errs most.
4. Inference phase: an uploaded tile flows through the trained model to produce a per-pixel class probability map; the highest-probability class wins each pixel, forming the land-cover mask.
5. The demo app overlays the predicted mask on the original tile with a color legend for each of the seven classes.
6. For full scenes, sliding-window tiled inference predicts overlapping windows and stitches them back into a seamless full-scene mask; masks export as PNG for report figures.

Specifications:
Model | U-Net with a pre-trained encoder (transfer learning)
Dataset | DeepGlobe Land Cover Classification (803 images, 7 classes)
Tile resolution | 2448x2448 pixels at 50 cm per pixel
Evaluation | Per-class IoU and mean Dice computed by the training notebook on the validation split
Design target | ~70%+ mean IoU (design target, not a measured claim)
Loss | Weighted cross-entropy + Dice loss for class imbalance
Inference | Sliding-window tiled prediction on CPU or GPU
Platform | Windows/Linux/macOS, Python 3.10; web demo app

[7-Class Land-Cover Segmentation] (implemented) — Classifies every pixel of a satellite tile into urban, agriculture, rangeland, forest, water, barren or unknown.
[U-Net with Pre-Trained Encoder] (implemented) — Encoder-decoder with skip connections preserves fine spatial detail; transfer learning from the pre-trained encoder speeds convergence on 803 images.
[Sliding-Window Tiled Inference] (implemented) — Overlapping tiles are predicted and stitched into seamless full-scene masks, handling arbitrarily large satellite scenes.
[Class-Imbalance Handling] (implemented) — Weighted cross-entropy plus Dice loss keeps rare classes like urban competitive against dominant barren and rangeland.
[Geospatial Augmentation] (implemented) — Rotations, flips and brightness jitter in the preprocessing pipeline improve robustness to orientation and illumination.
[Interactive Mask-Overlay Demo] (implemented) — The web app overlays predicted masks on uploaded tiles with a color legend; masks export as PNG for report figures.
[Per-Class IoU & Mean Dice Evaluation] (implemented) — The training notebook computes per-class IoU and mean Dice on the validation split with loss and IoU plots.
[Dataset Analysis & Tiling Scripts] (implemented) — Scripts handle dataset analysis, tiling, preprocessing and augmentation setup.

Complete source code (training, evaluation, inference, demo app)
Trained model weights (PyTorch .pt checkpoints)
Dataset preparation, tiling and augmentation scripts
Project report PDF (remote-sensing background, methodology, evaluation)
PPT presentation for the final review
Viva Q&A preparation document (U-Net, segmentation metrics, transfer learning, class imbalance)
Setup guide (environment, dependencies, dataset download steps)