About

This project uses a U-Net model to perform semantic segmentation on color infrared (CIR) imagery. The dataset belongs to the Geospatial Computing for Environmental Research (GCER) lab at MSState. It contains .tif raster images with embedded geospatial metadata, making it ideal for environmental monitoring and land cover classification.

The dataset is organized into three subsets:

datasets/
├── test
├── train
└── val

Model Architecture

U-Net is a popular CNN architecture designed for semantic segmentation tasks. Its symmetric encoder-decoder structure allows it to generate detailed segmentation masks, even with limited data. In this case, the model is trained on CIR imagery, which includes red, green, blue, and near-infrared bands. This helps improve the model’s ability to distinguish biophysical features (crops, forest, shrubs, water bodies, and so on) more accurately.

Known labels = ['Water', 'Raised Structures', 'Paved Structures', 'Barren Land', 'Forest', 'Herbacious', 'Cultivated Crops', 'Unclassified']

U-Net model architecture

The U-Net model simplified:

**Encoder (downsampling)**  
- One initial double conv layer:  (Conv2d → BatchNorm → ReLU) × 2  
- Four down blocks:  MaxPool → ((Conv2d → BatchNorm → ReLU) × 2)

**Bottleneck**  
- The last down block acts as the bottleneck

**Decoder (upsampling)**  
- Four up blocks:  ConvTranspose2d → pad → concat → ((Conv2d → BatchNorm → ReLU) × 2)

**Output**  
- 1×1 Conv2d layer

Example data

• Input:

• Ground truth label:

Training step

The model was trained on NVIDIA RTX4090 and training completed in under 30 mins using early stopping and a learning rate scheduler for efficient convergence.

• Segmentation during training:

• Training metrics:

• Confusion Matrix:

Final Evaluation

Reference: GCER