JavideuS/aid-image-classification

Model Card for Aerial Image Classification (CNN & Classic ML)

Model Details

Model Description

This repository contains two types of models for classifying aerial images from the AID dataset:

1. Convolutional Neural Network (CNN): A lightweight ResNet-based model.
2. Classic Machine Learning: A Bag of Features (BoF) pipeline using SIFT descriptors and Softmax Regression.

These models were developed as part of a machine learning assignment to evaluate deep learning approaches against classical computer vision methods.

• Model types:
  • CNN (PyTorch)
  • BoVW + ML algorithm (Scikit-learn/Joblib)
• Language(s): English
• Resources:
  • CNN: ~1M parameters, ~16MB.
  • Classic ML: ~100MB (includes vocabulary).

Model Architecture

The architecture consists of an initial convolution layer followed by three residual blocks.

• Residual Blocks: Enable deeper feature extraction without degradation.
• Layers: Convolution, Batch Normalization, Max-Pooling.
• Input: $600 \times 600$ pixel images (resized as needed by the pipeline).

Uses

Direct Use

The model is intended for classifying high-resolution aerial scenes into one of 30 categories. It is suitable for:

• Autonomous UAV navigation and mapping.
• Environmental monitoring.
• Land use classification.

Downstream Use

This model can be fine-tuned on other aerial or satellite imagery datasets.

Training Data

The model was trained on the Aerial Image Dataset (AID).

• Source: Google Earth imagery.
• Size: 10,000 images.
• Classes: 30 (e.g., Airport, Beach, Forest, Industrial, etc.).
• Split: 90% Training / 10% Test.

Performance

The CNN significantly outperformed classical Machine Learning methods (SVM, Random Forest, etc.) evaluated on the same dataset.

Metric	Value
Test Accuracy	92.80%
Macro Average	0.93
Weighted Average	0.93

Comparison with Classical Methods

Model	Test Accuracy
CNN (This Model)	0.9280
SVM (RBF Kernel)	0.7120
Softmax Regression	0.6580
Random Forest	0.5680
Naïve Bayes	0.5280

Limitations

• Data Bias: The model is trained on Google Earth imagery (AID), so it may not generalize perfectly to aerial images with significantly different sensors, resolutions, or lighting conditions.
• Scope: Limited to the 30 classes defined in the AID dataset.

How to Get Started

You can use the provided demo.ipynb notebook for a complete example. Below is a snippet to load both models.

1. Load Classic ML Model

from huggingface_hub import hf_hub_download
import joblib

# Download model
model_path = hf_hub_download(
    repo_id="JavideuS/aid-image-classification",
    filename="classicML/models/bovw_softmax.pkl"
)

# Load pipeline
bundle = joblib.load(model_path)
pipeline = bundle['pipeline']
label_encoder = bundle['label_encoder']

# Predict
# pipeline.predict(["path/to/image.jpg"])

2. Load CNN Model

import torch
from NeuralNets.model import PiattiCNN # Ensure you have the model definition
from huggingface_hub import hf_hub_download

# Download checkpoints
checkpoints_path = hf_hub_download(
    repo_id="JavideuS/aid-image-classification",
    filename="neuralNet/models/PiattiVL_v0.69.pth"
)

# Load model
checkpoints = torch.load(checkpoints_path, map_location='cpu')
model = PiattiCNN(num_classes=checkpoints['num_classes'])
model.load_state_dict(checkpoints['model_state_dict'])
model.eval()

# Inference
# ...

Citation

If you use this model or the AID dataset, please cite the original dataset paper:

@article{aid_dataset,
  title={AID: A Scene Classification Dataset},
  author={Xia, Gui-Song and et al.},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2017}
}