TLDR: A new research paper proposes a deep learning framework to estimate population in Gandhinagar, India, using high-resolution satellite imagery and Digital Elevation Models. The system employs a Convolutional Neural Network (CNN) to classify buildings as residential or non-residential with an F1-score of 0.9936, followed by an Artificial Neural Network (ANN) to estimate population at the building level. This automated method calculated Gandhinagar’s population at 278,954, offering a scalable and efficient alternative to traditional census methods for urban planning and resource allocation.
Estimating population accurately is vital for effective urban planning, resource distribution, and emergency response. Traditional methods like surveys and censuses are often costly, time-consuming, and require significant human effort, struggling to keep pace with the rapid changes in modern urban environments.
A new study introduces an innovative deep learning solution to overcome these challenges, focusing on the urban area of Gandhinagar, India. This approach leverages high-resolution satellite imagery (0.3m), Digital Elevation Models (DEM) of 0.5m resolution, and vector boundaries to provide a more efficient and precise way to count people.
A Dual Deep Learning Approach
The core of this research is a dual-model framework that combines two types of neural networks: a Convolutional Neural Network (CNN) and an Artificial Neural Network (ANN). The CNN is tasked with a crucial initial step: classifying buildings as either residential or non-residential. This distinction is fundamental because only residential buildings contribute to the population count.
The CNN model was trained using approximately 48,000 building footprints from Gandhinagar. It processes small image patches (64×64 pixels) centered on each building, learning to identify characteristics that differentiate residential from non-residential structures. The model demonstrated remarkable accuracy, achieving an overall F1-score of 0.9936, indicating its strong ability to correctly classify buildings.
Once buildings are classified, the ANN takes over. For residential buildings, it estimates the population. Since actual ground-truth population data for individual buildings is often unavailable, the researchers developed a rule-based function to generate ‘synthetic’ population labels for training the ANN. This function estimates population based on building area and height, assuming a certain number of residents per floor depending on the building’s size.
The ANN then learns from these synthetic labels, using features like the mean pixel intensity of the building’s image patch, its area in square feet, and its height in meters to predict the number of occupants. This regression-based approach allows for a direct prediction of population at the building level.
Gandhinagar’s Population Through AI
The study focused on Gandhinagar, the capital of Gujarat, India, a planned city known for its organized layout. High-resolution imagery from the WorldView-3 satellite, acquired in November 2023, along with a DEM generated from stereo pairs, provided the detailed spatial data needed for this analysis.
Using this advanced framework, the study estimated Gandhinagar’s urban population at 278,954. This figure aligns closely with census-projected estimates for 2025, which anticipate the city’s population to be around 300,000. The consistency of these results with official projections highlights the reliability and accuracy of the proposed deep learning method.
Also Read:
- DGMap: A New Approach to Road Network Inference
- Advancing Open-Vocabulary Segmentation for Remote Sensing Images
Implications for Urban Governance
This automated approach addresses significant limitations of conventional census-based methodologies. By integrating real-time data updates, standardized metrics, and infrastructure planning capabilities, it offers municipalities a scalable and replicable tool for optimized resource management in rapidly urbanizing cities. The research showcases the efficiency of AI-driven geospatial analytics in enhancing data-driven urban governance.
While the model shows great promise, the authors acknowledge certain limitations. These include potential biases from class imbalance in the training dataset and the assumption of uniform occupancy rates, which might not always reflect the varied uses of buildings. Future work will involve verifying this methodology in other urban areas and conducting ground surveys to further enhance accuracy.
For more details, you can read the full research paper here.
