bhargav37/lulc-dl-model

🛰️ Land Cover Segmentation — U-Net

A U-Net deep learning model for pixel-wise land cover classification from Sentinel-2 multi-spectral satellite imagery. Trained to identify 5 land cover types across large-scale geospatial datasets.

---

🗺️ Land Cover Classes

The model is trained to classify every pixel into one of the following 5 categories:

Class	Description
Barren Land	Exposed soil, rock, and sparsely vegetated areas
Built-up Area	Urban and suburban structures, roads
Crop	Agricultural and cultivated farmland
Forest	Dense tree cover and woodland areas
Water	Rivers, lakes, reservoirs, and water bodies

---

🏗️ Model Architecture

The model is based on the U-Net architecture — a fully convolutional encoder-decoder network with skip connections designed for semantic image segmentation.

Encoder progressively extracts spatial features through two blocks of convolutions followed by max pooling, reducing the spatial resolution while increasing feature depth (32 → 64 filters).

Bottleneck captures the highest-level abstract features at the compressed representation (128 filters).

Decoder restores the original spatial resolution through two upsampling blocks. At each step, skip connections from the corresponding encoder block are concatenated to recover fine-grained spatial detail (64 → 32 filters).

Output is a 1×1 convolution with softmax activation that produces a per-pixel probability distribution over the 5 land cover classes.

Input (64×64×16)
    │
    ├─── Encoder Block 1 — Conv(32) → Conv(32) → MaxPool ──────────────┐ skip
    ├─── Encoder Block 2 — Conv(64) → Conv(64) → MaxPool ────────┐ skip│
    ├─── Bottleneck      — Conv(128) → Conv(128)                  │     │
    ├─── Decoder Block 1 — UpSample → Concat ←───────────────────┘     │
    │                      Conv(64) → Conv(64)                          │
    ├─── Decoder Block 2 — UpSample → Concat ←─────────────────────────┘
    │                      Conv(32) → Conv(32)
    └─── Output          — Conv(5, 1×1) → Softmax
                           (64×64×5)

---

📥 Input

Property	Details
Sensor	Sentinel-2 Multi-Spectral Imagery
Patch Size	64 × 64 pixels
Channels	16 spectral bands
Preprocessing	Normalized pixel values

---

📤 Output

Property	Details
Shape	64 × 64 pixel segmentation mask
Type	Per-pixel class label (one of 5 land cover classes)
Format	Integer class map derived from softmax probabilities

---

📊 Model Performance

Evaluated on 5,746,688 test pixels across all 5 classes.

Overall Metrics

Metric	Score
Overall Accuracy	93.04%
Validation Accuracy	93.04%
Validation Loss	0.2678
Macro Avg Precision	92.58%
Macro Avg Recall	92.88%
Macro Avg F1-Score	92.69%
Weighted Avg F1	93.06%

Per-Class Performance

Class	Precision	Recall	F1-Score
Barren Land	84.30%	90.59%	87.33%
Built-up Area	93.32%	89.29%	91.26%
Crop	93.53%	95.42%	94.47%
Forest	95.55%	96.84%	96.19%
Water	96.23%	92.27%	94.21%

Forest and Water achieve the highest classification accuracy. Barren Land is the most challenging class, likely due to spectral overlap with Built-up Areas and Crop fields.

---

📁 Model File

File	Description
landcover_unet_model.h5	Trained Keras model (weights + architecture)