LandUseVision

LandUseVision is a computer vision project that classifies land-use images into four categories:

• Forest
• Industrial
• Permanent Crop
• Residential

The project uses transfer learning with a pretrained ResNet18 model from PyTorch/Torchvision. The final classification layer is adapted for the four project classes, then trained on the included image dataset.

Project Goal

The goal of this project is to build an image classifier that can recognize different land-use categories from image data. This type of classification is commonly used in remote sensing, urban planning, agriculture, and environmental monitoring.

Repository Contents

LandUseVision/
  DataSet/
    Train/
    Validate/
    Test/
  results/
  model_utils.py
  train.py
  main.py
  evaluate.py
  predict.py
  demo.py
  model.pth
  requirements.txt
  LICENSE

Dataset

The dataset is organized into training, validation, and test folders. Each folder contains class-specific subfolders.

DataSet/
  Train/
    Forest/
    Industrial/
    PermanentCrop/
    Residential/
  Validate/
    ForestValidate/
    IndustrialValidate/
    PermanentCropValidate/
    ResidentialValidate/
  Test/
    ForestTest/
    IndustrialTest/
    PermanentCropTest/
    ResidentialTest/

The current dataset contains approximately:

Split	Forest	Industrial	Permanent Crop	Residential
Train	2400	2000	2000	2400
Validate	300	250	250	300
Test	300	250	250	300

Model

The training script uses a pretrained ResNet18 model and replaces the final fully connected layer with a classifier for the four land-use classes.

Image preprocessing includes:

• Resizing images to 224 x 224
• Converting images to tensors
• Normalizing images with ImageNet mean and standard deviation values

Results

The included model.pth checkpoint achieves the following results on the test set:

Metric	Result
Overall accuracy	98.64%

Per-class accuracy:

Class	Accuracy	Correct / Total
Forest	100.00%	300 / 300
Industrial	97.60%	244 / 250
Permanent Crop	97.20%	243 / 250
Residential	99.33%	298 / 300

Example single-image prediction:

Prediction: Residential
Confidence: 100.00%

Evaluation and training plots are saved in the results/ directory by default.

Setup

This project uses Python with PyTorch and Torchvision. Python 3.10 or newer is recommended.

Clone the repository and move into the project folder:

git clone https://github.com/IcebergSnow/LandUseVision.git
cd LandUseVision

Create and activate a virtual environment:

python -m venv .venv
source .venv/bin/activate

On Windows, activate the virtual environment with:

.venv\Scripts\activate

Install the required packages:

pip install -r requirements.txt

How to Run

The repository includes a saved model file, model.pth, so you can evaluate the model or run predictions without training it first.

Run the demo:

python demo.py

The demo evaluates the saved model, saves results/confusion_matrix.png, and predicts the class for one sample image.

Evaluate the saved model on the test set:

python evaluate.py

The evaluation script prints overall accuracy, per-class accuracy, and saves a confusion matrix image to results/confusion_matrix.png.

Predict the class for one image:

python predict.py DataSet/Test/ResidentialTest/Residential_332.jpg

The prediction script prints the predicted class and confidence score.

Run the demo with a different image:

python demo.py --image-path DataSet/Test/ForestTest/Forest_301.jpg

Skip evaluation and only run the sample prediction:

python demo.py --skip-evaluation

Train the model again:

python train.py

The training script reports training loss, validation loss, and validation accuracy for each epoch. It saves the best model weights to model_weights.pth based on validation accuracy and saves a training curve plot to results/training_curves.png.

Custom output paths can be provided with --confusion-matrix-path for evaluation and --metrics-plot-path for training.

For compatibility, python main.py also starts training.

All scripts expect the dataset to be located in the DataSet/ folder using the structure shown above.

Checkpoints

New training runs save model weights to model_weights.pth, which is the recommended PyTorch checkpoint format for this project.

Evaluation and prediction prefer model_weights.pth when it exists. If it does not exist, they fall back to the existing full-model checkpoint, model.pth, so the project still works immediately after cloning.

Device Note

The training, evaluation, and prediction scripts automatically choose between Apple Silicon mps, NVIDIA cuda, and cpu depending on what is available.

License

This project is licensed under the MIT License. See LICENSE for details.