Height Segmentation Model for Urban Planning Using Satellite Imagery

Abstract

The "Height Segmentation Model" is a deep learning-driven project aimed at analyzing satellite imagery to estimate building heights and generate accurate segmentation maps for urban environments. By leveraging advanced techniques in convolutional neural networks (CNNs) and utilizing a custom-built ResUNet architecture, this model provides both building segmentation and height estimation as key outputs. The project involves the entire pipeline from data preprocessing to model deployment, including a Streamlit web application to visualize results interactively.

Introduction

Urban planning and development benefit greatly from the availability of accurate and detailed data. Satellite imagery offers a non-intrusive way to collect geographic data, which, when paired with deep learning, can provide high-resolution segmentation maps of cities, identifying buildings and estimating their heights. This model focuses on segmentation of satellite images to distinguish between buildings and background, enabling efficient urban analysis for applications like city planning, disaster management, and infrastructure development.

Key Features

• Height Estimation: The model predicts the height of buildings based on satellite images.
• Building Segmentation: A binary mask is generated, highlighting buildings in the satellite images.
• Streamlit App: An interactive web interface that allows users to upload satellite images, visualize the segmentation, and download results.
• Performance Insights: Displays key performance metrics like Accuracy, IoU, and F1-Score for model evaluation.

Methodology

1. Data Collection and Preprocessing:

The project begins by collecting high-resolution satellite images and corresponding ground truth labels for training. The images undergo preprocessing steps such as resizing, normalization, and augmentation to prepare them for model input.

2. Model Architecture:

The ResUNet model, an extension of the U-Net architecture, is used for building segmentation. This model is particularly suited for segmentation tasks because of its skip connections and deep learning capabilities. The model architecture comprises an encoder-decoder structure with residual connections to preserve important features throughout the network layers.

3. Model Training:

Using the preprocessed images and their corresponding masks, the ResUNet model is trained with a combination of binary cross-entropy loss and IoU (Intersection over Union) as the evaluation metric. Training is performed on a GPU to speed up the process, and model checkpoints are saved throughout.

4. Deployment:

After training, the model is integrated into a Streamlit application. This app allows users to interact with the model in real-time, uploading satellite images to generate building segmentation masks. The application also displays performance metrics and allows for the downloading of results.

Results

Sample Input and Output

Input Satellite Image:

Output Segmentation Map:

Performance Metrics

• Accuracy: 92%
• IoU: 0.85
• F1-Score: 0.89

Insights

The model provides detailed performance metrics through the Streamlit interface, enabling easy evaluation and comparison of different models.

Future Enhancements

• 3D Building Modeling: In future iterations, we plan to incorporate 3D modeling techniques to enhance visualizations and provide more accurate building height predictions.
• Vegetation and Urban Feature Analysis: The model could be expanded to analyze other urban features such as vegetation and roads.
• Optimized Inference Speed: Reducing the inference time for larger datasets will be a key focus in the next phase of development.
• Deployment on Cloud Platforms: Future updates include deploying the application on cloud platforms like AWS, GCP, or Streamlit Community Cloud to make it accessible worldwide.

Team Members

The team behind the Height Segmentation Model includes passionate data science interns from the ExploreAI Academy. They are:

• Emmanuel Nkhubalale - Email
• Nolwazi Mndebele - Email
• Carroll Tshabane - Email
• Jan Motene - Email
• Muwanwa Tshikovhi - Email
• Sibukiso Nhlengethwa - Email
• Zamancwabe Makhathini - Email

How to Use the Model

1. Clone the Repository:
   Clone this repository to your local machine:

   git clone https://github.com/username/segmentation-model-app.git
   cd segmentation-model-app
   

2. Install Dependencies: Install the necessary libraries by running:

   undefined

3. Prepare the Model: Place the trained model (final_model_resunet_finetuned_lr_reduced_5.h5) in the root directory.

4. Run the App: To start the Streamlit application, use the following command:

streamlit run app.py

## Acknowledgments

We would like to acknowledge the following for their support and contributions to the project:

Hugging Face for providing access to high-quality datasets.
ExploreAI Academy for fostering innovation in data science.
Streamlit for providing an easy-to-use platform for interactive app development.

## Conclusion
The Height Segmentation Model leverages deep learning and satellite imagery to solve a critical problem in urban planning. By providing building height estimates and segmentation maps, the project paves the way for future applications in smart city development and resource management.

vbnet
Copy code