Abstract
The global population at risk of mosquito-borne diseases such as dengue, yellow fever, chikungunya and Zika is expanding. Infectious disease models commonly incorporate environmental measures like temperature and precipitation. Given increasing availability of high-resolution satellite imagery, here we consider including landscape features from satellite imagery into infectious disease prediction models. To do so, we implement a Convolutional Neural Network (CNN) model trained on Imagenet data and labelled landscape features in satellite data from London. We then incorporate landscape features from satellite image data from Pakistan, labelled using the CNN, in a well-known Susceptible-Infectious-Recovered epidemic model, alongside dengue case data from 2012-2016 in Pakistan. We study improvement of the prediction model for each of the individual landscape features, and assess the feasibility of using image labels from a different place. We find that incorporating satellite-derived landscape features can improve prediction of outbreaks, which is important for proactive and strategic surveillance and control programmes.
Introduction
Predicting the spread of vector-borne diseases like dengue is notoriously difficult. Traditional epidemiological models typically rely on "coarse" variables such as temperature, precipitation, and general population density. While these factors influence mosquito life cycles, they often ignore the micro-environments—specific building clusters or drainage patterns—where mosquitoes actually breed and interact with humans.
While high-resolution satellite imagery (e.g., Google Maps) is increasingly available, manually labeling these features across entire countries is impractical. Furthermore, many developing regions lack the localized training data needed to build region-specific AI models. This study explores the use of transfer learning to extract meaningful landscape data from one geography (London) to predict disease outbreaks in another (Pakistan).
Methodology
The research utilizes a multi-step pipeline that integrates computer vision with classical epidemiology. The process begins with raw satellite imagery and concludes with a refined disease prediction model.
Semantic Segmentation and Transfer Learning
To identify landscape features, the authors employed a U-Net architecture with a VGG16 encoder pre-trained on ImageNet. To address the lack of labeled data in Pakistan, the decoder was fine-tuned using the Kaggle DSTL dataset, which contains high-resolution imagery of London. The model was trained to identify six landscape classes: buildings, roads, trees, crops, waterways, and standing water.
Feature Extraction and Integration
The trained model was applied to Google Maps API imagery of Lahore and Rawalpindi, Pakistan (~0.6m resolution). This converted raw pixels into "area percentages" for each landscape feature within specific sub-city units. These percentages were then integrated into a Timeseries Susceptible-Infectious-Recovered (TSIR) model. In this modified model, landscape features act as time-invariant factors influencing the transmission rate ($\beta$).
Key Results
The inclusion of high-resolution landscape features significantly improved the predictive power of the epidemiological models compared to base models using only environmental and population data.
- Model Fit Improvement: The base model (weather + population) achieved an adjusted $R^2$ of 0.728 for Lahore and 0.774 for Rawalpindi. Incorporating all landscape features boosted these scores to 0.747 and 0.808, respectively.
- Feature Impact: Buildings and waterways provided the most significant improvements to predictive accuracy.
- Urban Signal: The model showed a stronger performance boost in highly urbanized areas, suggesting that man-made landscapes are dominant drivers of disease in dense cities.
Discussion and Conclusion
A key finding of this work is the transferability of structural features. While biological features like "crops" or "standing water" were difficult to generalize across different climates, structural features like buildings and roads transferred effectively from London to Pakistan. This proves the viability of cross-continental transfer learning for public health applications.
Future work may involve using multi-temporal imagery to track urban changes in real-time and exploring unsupervised feature extraction to bypass the labeling bottleneck entirely. By moving to sub-meter resolution, researchers can now capture the primary scales at which mosquito ecology operates, enabling more proactive and strategic disease control programmes.
Summary updated