A cloud-aware spatio-temporal framework for analyzing historical urban expansion and forecasting future building development from multi-temporal satellite imagery.
Urban areas evolve continuously, making timely monitoring and forecasting important for urban planning, infrastructure development, and environmental management. Traditional satellite-image analysis often relies on manual interpretation, while many deep learning approaches focus on individual images without explicitly modelling how urban development changes over time.
This project presents an end-to-end system for historical urban growth analysis and future urban expansion prediction using multi-temporal satellite imagery from the SpaceNet-7 dataset.
The framework combines:
- K-Means clustering and spatial analysis for historical urban-growth analysis
- CNNs for spatial feature extraction
- Transformers for long-range temporal dependency modelling
- LSTMs for sequential urban-growth modelling
- U-Net with a ResNet-50 encoder for cloud segmentation
- Auto-regressive forecasting for multi-step future prediction
- FastAPI backend for model inference and analysis
- HTML/CSS/JavaScript frontend for interactive visualization
The proposed cloud-aware hybrid model achieved an IoU of 0.62, with a Dice score of 0.74, Precision of 0.76, Recall of 0.68, and F1-score of 0.72 on the evaluated test data.
Analyze changes across a temporal sequence of satellite images and extract:
- Baseline building count
- Final building count
- Newly developed buildings
- Baseline built-up area
- Final built-up area
- Monthly area change
- Cumulative urban growth
- Seasonal development trends
- Spatial development patterns
The predictive framework combines three complementary deep learning components:
CNN → Transformer → LSTM
- CNN extracts spatial features from satellite imagery.
- Transformer models relationships between observations across time.
- LSTM captures sequential patterns in urban development.
- A prediction layer generates future building masks.
Optical satellite imagery can contain cloud-covered regions where buildings and land-use information become unreliable.
To address this, the framework incorporates a cloud segmentation module based on:
U-Net + ResNet-50 encoder
The module generates pixel-level cloud masks that identify unreliable regions before prediction.
The system uses an auto-regressive forecasting strategy in which predicted building masks can be fed back into subsequent prediction steps, allowing the model to generate forecasts across multiple future time steps.
The project includes a browser-based interface that allows users to:
- Upload a temporal stack of
.tif/.tiffsatellite images - Run the analysis pipeline
- Explore historical urban development
- View monthly and seasonal trends
- View predicted building masks
- Visualize predicted changes
- Inspect cloud masks and overlays
- View urban-growth and building-growth classifications
Multi-Temporal Satellite Images
│
▼
┌─────────────────────┐
│ Data Preprocessing │
│ │
│ • CRS verification │
│ • Resizing │
│ • RGB validation │
│ • Normalization │
│ • Spatial alignment │
└──────────┬──────────┘
│
┌─────────────┴─────────────┐
│ │
▼ ▼
┌──────────────────┐ ┌──────────────────────┐
│ Historical │ │ Cloud Segmentation │
│ Urban Analysis │ │ │
│ │ │ U-Net + ResNet-50 │
│ K-Means │ │ │
│ Spatial Analysis │ │ → Cloud Mask │
│ Monthly Trends │ └──────────┬───────────┘
│ Seasonal Trends │ │
└────────┬─────────┘ │
│ ┌──────────▼───────────┐
│ │ CNN Spatial Features │
│ └──────────┬───────────┘
│ │
│ ┌──────────▼───────────┐
│ │ Transformer │
│ │ Temporal Attention │
│ └──────────┬───────────┘
│ │
│ ┌──────────▼───────────┐
│ │ LSTM │
│ │ Sequential Growth │
│ └──────────┬───────────┘
│ │
│ ┌──────────▼───────────┐
│ │ Predicted Building │
│ │ Mask │
│ └──────────┬───────────┘
│ │
└──────────────┬──────────────┘
▼
Urban Growth Visualization
and Development Classification
The proposed cloud-aware hybrid pipeline combines cloud segmentation, CNN-based spatial feature extraction, Transformer-based temporal attention, LSTM-based sequential modelling, and auto-regressive prediction.
The historical analysis captures spatial urban-density patterns and changes in development across the temporal observation period.
The system tracks month-by-month building development and urban-area change to identify temporal growth patterns.
The predictive model generates future building masks that can be compared with corresponding ground-truth building footprints.
Predicted building masks can be iteratively fed back into the model to generate multi-step future urban-growth forecasts.
The cloud-aware component identifies cloud-contaminated regions and supports improved building delineation in areas affected by cloud cover.
Training curves show the progression of loss, accuracy, and IoU during model training.
The project uses the SpaceNet-7 multi-temporal satellite imagery dataset.
SpaceNet-7 provides high-resolution satellite imagery and building-footprint annotations for studying urban development and change over time.
- Multi-temporal satellite imagery
- More than 60 urban locations
- Up to 24 monthly observations per location
- Observation period covering approximately 2018–2020
- RGB GeoTIFF satellite imagery
- Building footprints provided as GeoJSON annotations
- Ground sampling distance of approximately 30 cm
The building footprints serve as ground-truth annotations for supervised training and evaluation.
Note: The SpaceNet-7 dataset is not included in this repository because of its size and dataset-specific distribution terms. Download and prepare the dataset separately before running the complete pipeline.
Before analysis and model inference, the satellite imagery undergoes several preprocessing steps.
Geospatial metadata is checked using Rasterio to ensure that imagery is correctly georeferenced and spatially aligned.
Satellite tiles are resized to:
256 × 256 pixels
Input images are verified to contain three RGB channels.
Temporal images are checked to ensure that their spatial footprint remains consistent across observations.
Pixel values are scaled to:
[0, 1]
GeoJSON building annotations are rasterized and aligned with the corresponding image tiles to generate binary building masks for supervised learning and evaluation.
The research pipeline uses a non-overlapping:
| Split | Percentage |
|---|---|
| Training | 70% |
| Validation | 15% |
| Testing | 15% |
The first stage focuses on understanding how urban development occurred historically.
Building-footprint spatial information is analyzed using K-Means clustering.
Latitude and longitude information from building polygons is used to identify spatial patterns of urban development over different time steps.
The analysis provides:
- Monthly development patterns
- Spatial density patterns
- Changes in building counts
- Changes in built-up area
- Cumulative development trends
- Seasonal construction patterns
For consecutive observations, the system calculates:
New Buildings = Current Building Count - Previous Building Count
and tracks the corresponding change in built-up area.
Monthly development observations are aggregated into seasonal summaries to identify variations in construction activity.
The analysis can be visualized through:
- Urban-density heatmaps
- Monthly development charts
- Area-change charts
- Cumulative growth trends
- Seasonal analysis
- Monthly records table
The predictive component uses a hybrid architecture combining spatial and temporal learning.
The CNN processes satellite imagery and extracts spatial representations of:
- Buildings
- Roads
- Urban structures
- Local spatial patterns
The cloud information is incorporated into the predictive pipeline so that cloud-affected regions can be identified.
The Transformer processes the sequence of spatial features and uses self-attention to model relationships across different observations in the temporal sequence.
This allows the model to capture long-range temporal dependencies that may not be fully represented by sequential processing alone.
The Transformer output is passed to LSTM layers.
The LSTM models the step-by-step evolution of urban development, learning how construction activity changes across the observation period.
The final prediction stage produces a binary building mask indicating regions where future development is expected.
Cloud contamination is a major challenge when working with optical satellite imagery.
Cloud-covered regions may hide buildings and introduce unreliable information into the prediction pipeline.
A U-Net-based segmentation module with a ResNet-50 encoder is used to identify cloud-covered regions.
Satellite Image
│
▼
U-Net + ResNet-50
│
▼
Binary Cloud Mask
│
▼
Cloud-Aware Prediction Pipeline
The resulting cloud mask explicitly identifies unreliable image regions.
The cloud-aware architecture therefore combines:
CNN
+
Transformer
+
LSTM
+
Cloud Segmentation
To extend prediction beyond a single future time step, the predicted building mask can be fed back into the model as part of the next input sequence.
Historical Images
│
▼
Prediction t+1
│
▼
Feed prediction back
│
▼
Prediction t+2
│
▼
Feed prediction back
│
▼
Prediction t+3
│
▼
...
This allows the framework to generate multi-step forecasts of future urban development.
Longer forecasting horizons can introduce accumulated prediction error, which is an important consideration when interpreting multi-step forecasts.
The research model was trained using:
- Loss: Binary Cross-Entropy
- Optimizer: Adam
- Learning Rate: 0.001
- Input: Multi-temporal satellite imagery
- Target: Binary building masks
The predictive model is evaluated against ground-truth building masks using:
Measures the overlap between the predicted and ground-truth building regions.
Measures similarity between predicted and ground-truth masks.
Measures the proportion of predicted building pixels that correspond to actual building pixels.
Measures the proportion of actual building pixels correctly detected.
Combines precision and recall into a single metric.
| Metric | Score |
|---|---|
| IoU | 0.62 |
| Dice | 0.74 |
| Precision | 0.76 |
| Recall | 0.68 |
| F1-Score | 0.72 |
| Model | IoU | Dice | Precision | Recall | F1 |
|---|---|---|---|---|---|
| FCN | 0.42 | 0.55 | 0.60 | 0.50 | 0.54 |
| SegNet | 0.45 | 0.58 | 0.63 | 0.52 | 0.57 |
| U-Net | 0.50 | 0.65 | 0.69 | 0.58 | 0.63 |
| ConvLSTM | 0.53 | 0.68 | 0.71 | 0.60 | 0.65 |
| Transformer | 0.54 | 0.69 | 0.72 | 0.61 | 0.66 |
| Proposed | 0.62 | 0.74 | 0.76 | 0.68 | 0.72 |
The proposed model achieved the highest IoU, Dice, Precision, Recall, and F1-score among the compared approaches.
The contribution of individual architectural components was evaluated through an ablation study.
| Model Variant | IoU | Dice | Precision | Recall | F1 |
|---|---|---|---|---|---|
| CNN Only | 0.44 | 0.57 | 0.61 | 0.49 | 0.54 |
| CNN + LSTM | 0.49 | 0.63 | 0.67 | 0.55 | 0.60 |
| CNN + Transformer | 0.52 | 0.66 | 0.70 | 0.58 | 0.63 |
| Hybrid CNN + Transformer + LSTM | 0.56 | 0.71 | 0.73 | 0.65 | 0.69 |
| Hybrid + Cloud Awareness | 0.62 | 0.74 | 0.76 | 0.68 | 0.72 |
The progression demonstrates the contribution of the individual components:
CNN
↓
CNN + LSTM
↓
CNN + Transformer
↓
CNN + Transformer + LSTM
↓
Cloud-Aware Hybrid
The IoU increased from 0.44 for the CNN-only model to 0.62 for the final cloud-aware hybrid framework.
The cloud-aware extension further improved IoU from 0.56 to 0.62, indicating the value of explicitly identifying cloud-contaminated regions.
The project includes a web-based interface for interacting with the analysis pipeline.
Upload temporal satellite-image stack
│
▼
Run Analysis
│
┌───────┴────────┐
│ │
▼ ▼
Phase 1 Phase 2
Historical Prediction
Analysis │
│ ▼
│ Predicted Mask
│
└────────┬──────────┐
│ │
▼ ▼
Phase 3 Cloud Analysis
│
▼
Urban Growth Classification
The frontend displays:
- Baseline area
- Final area
- Total new buildings
- Total area change
- Baseline building count
- Final building count
- Monthly building development
- Monthly area change
- Cumulative urban change
- Seasonal analysis
- Monthly records table
The prediction interface displays:
- Last image in the input stack
- Predicted building mask
- Highlighted predicted changes
The classification and cloud-analysis interface displays:
- Urban development classification
- Building growth classification
- Observed development
- Predicted development
- Growth percentage
- Predicted area
- Predicted buildings
- Cloud mask
- Cloud overlay
- Python
- TensorFlow / Keras
- NumPy
- SciPy
- scikit-learn
- Matplotlib
- Pillow
- Rasterio
- OpenCV
- CNN
- Transformer
- LSTM
- U-Net
- ResNet-50
- FastAPI
- Uvicorn
- Python Multipart
- HTML5
- CSS3
- JavaScript
- Chart.js
Urban-Development-Analysis-and-Prediction/
│
├── backend/
│ ├── main.py
│ ├── analysis.py
│ ├── model_utils.py
│ │
│ └── models/
│ └── model.h5
│
├── frontend/
│ ├── index.html
│ ├── style.css
│ └── app.js
│
├── requirements.txt
├── .gitignore
└── README.md
main.py
FastAPI application responsible for:
- Receiving uploaded satellite images
- Running the analysis pipeline
- Loading the model
- Running prediction
- Running cloud segmentation
- Generating classification results
- Returning visualization data
analysis.py
Contains the historical urban-growth analysis functionality, including:
- Building-count estimation
- Built-up area estimation
- Monthly development analysis
- Seasonal aggregation
model_utils.py
Contains:
- Model loading
- Image preprocessing
- Prediction
- Cloud segmentation
- Visualization generation
- Urban-growth classification
index.html
Defines the user interface and result sections.
app.js
Handles:
- Satellite-image upload
- Temporal ordering
- API communication
- Result rendering
- Chart generation
- Prediction visualization
style.css
Defines the application's user interface and visualization styling.
git clone https://github.com/PavanKumarBaduru/Urban-Development-Analysis-and-Prediction.git
cd Urban-Development-Analysis-and-Predictionpython -m venv venv
venv\Scripts\activatepython3 -m venv venv
source venv/bin/activatepip install -r requirements.txtDownload and prepare the SpaceNet-7 dataset separately.
The application expects a temporal stack of RGB satellite images in:
.tif
or
.tiff
format.
The application sorts uploaded files alphabetically to preserve temporal ordering.
Therefore, filenames should follow a chronological naming convention, for example:
2019_01.tif
2019_02.tif
2019_03.tif
...
2020_01.tif
Incorrect filename ordering can result in an incorrect temporal sequence.
From the backend directory:
cd backend
uvicorn main:app --reloadThe API will run locally on:
http://127.0.0.1:8000
The backend provides a health endpoint:
GET /health
A successful response is:
{
"status": "ok"
}The main analysis endpoint is:
POST /analyze
It accepts multiple satellite-image files.
At least two .tif/.tiff images are required for the temporal analysis pipeline.
After starting the backend, open:
frontend/index.html
in a browser.
The frontend communicates with the FastAPI backend at:
http://127.0.0.1:8000
Select multiple temporally ordered satellite images and click:
Run Analysis
The resulting Phase 1, Phase 2, and Phase 3 outputs are then displayed through the web interface.
The project has several limitations identified during evaluation:
- Optical satellite imagery can still be affected by persistent cloud cover.
- Long forecasting horizons can accumulate prediction errors through auto-regressive feedback.
- The hybrid architecture has higher computational requirements than simpler segmentation models.
- Historical clustering provides a useful representation of spatial development patterns but does not capture every fine-grained structural change.
- Performance can degrade when usable imagery is sparse for extended periods.
- The current framework primarily uses RGB optical imagery.
Potential extensions include:
- Longer forecasting horizons
- Multi-spectral satellite imagery
- Synthetic Aperture Radar (SAR) data for improved cloud robustness
- More advanced cloud-aware feature fusion
- Incorporation of socio-economic factors
- Improved high-resolution urban-growth modelling
- More robust multi-step forecasting
- Deployment of the complete inference system as a scalable cloud service
The project demonstrates an end-to-end approach to urban development monitoring by combining historical spatial analysis, deep spatio-temporal modelling, cloud segmentation, and future-growth forecasting.
The major contributions are:
- Historical urban-growth analysis using clustering and spatial information.
- Hybrid CNN–Transformer–LSTM modelling for spatio-temporal urban expansion prediction.
- Cloud-aware prediction using U-Net and a ResNet-50 encoder.
- Auto-regressive forecasting for multi-step future development prediction.
- Interactive FastAPI-based application for analyzing satellite-image sequences and visualizing predictions.
- Comparative and ablation analysis demonstrating the contribution of the proposed architecture.
The final cloud-aware model achieved:
IoU: 0.62 | Dice: 0.74 | Precision: 0.76 | Recall: 0.68 | F1: 0.72
Detailed project documentation will be added here:
A project license has not yet been specified.
Research Prototype / Academic Project