Deep Landscape Features for Improving Vector-borne Disease Prediction

Nabeel Abdur Rehman, Umar Saif, Rumi Chunara

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

Vector-borne diseases like dengue, Zika, and malaria are a massive global health burden, causing over 700,000 deaths annually. Traditionally, epidemiologists predict outbreaks using environmental variables like temperature and precipitation, as these affect mosquito life cycles. However, these "coarse" climate models often miss the local reality on the ground.

Mosquitoes require specific habitats—such as standing water in construction sites, shaded areas under trees, or high-density urban corridors—to spread disease effectively. While high-resolution satellite imagery exists, manually labeling every building and puddle across an entire country is impossible. This research addresses the challenge of automatically extracting these "micro-features" of the landscape and demonstrating their utility in improving disease prediction models.

Methodology

The research employs a multi-step pipeline combining computer vision with classical epidemiology to bridge the gap between satellite data and disease forecasting.

Overview of the research pipeline, from satellite image segmentation to time-series epidemic modeling.

Semantic Segmentation and Feature Extraction

A U-Net architecture was utilized for pixel-level classification. To address the lack of labeled data for Pakistan, the researchers used transfer learning: a VGG16 model pre-trained on ImageNet was combined with a decoder trained on the Kaggle DSTL dataset (London-based imagery). The model identified six landscape classes: Buildings, Roads, Trees, Crops, Waterways, and Standing Water.

Epidemic Modeling

Extracted features were converted into area proportions for different city sectors and integrated into a Time-series Susceptible-Infected-Recovered (TSIR) model. This model accounts for the number of susceptible and infected individuals, using the landscape features (e.g., urban density) to influence the transmission rate ($\beta$).

Datasets

  • Satellite Imagery: High-resolution images from Google Maps API (zoom level 17) covering Lahore and Rawalpindi, Pakistan.
  • Epidemiological Data: Confirmed dengue case data (2012–2017) from the Punjab Information Technology Board.
  • Environmental Data: Daily temperature and precipitation from the Pakistan Meteorological Department and population density from WorldPop.

Results and Discussion

The integration of landscape features significantly improved model accuracy, as measured by the Adjusted $R^2$ metric. The most substantial gains were observed when incorporating data on buildings and waterways, suggesting that the built environment is a primary driver of human-mosquito interaction in these regions.

Comparison of model performance (Adjusted R²) across different landscape features and urban densities in Lahore and Rawalpindi.

Key Insights

  • The "Urban Signal": In Lahore, the model's fit improved from 0.728 (climate only) to 0.747 (climate + landscape). In Rawalpindi, the improvement was even more pronounced, jumping from 0.774 to 0.808.
  • Cross-Continental Transfer Learning: A surprising finding was that a model trained on London labels could successfully identify urban features in Pakistan. This indicates that certain deep features of urbanization are universal enough to be applied in data-poor regions.
  • Limitations of Natural Feature Transfer: The model struggled with "Crops" and "Standing Water," likely because these features look visually distinct in Pakistan compared to the London training set. This highlights the limits of cross-geographic transfer learning for natural versus man-made features.

Conclusion

Incorporating satellite-derived landscape features into epidemic models provides a more granular and accurate prediction of disease outbreaks than climate data alone. While the study used static landscape snapshots, future work could involve real-time tracking of environmental changes using higher-frequency satellite passes. These findings are critical for developing proactive and strategic surveillance and control programs in regions vulnerable to vector-borne diseases.

Summary updated

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.
BibTeX
@inproceedings{rehman2019deep,
  author = {Nabeel Abdur Rehman and Umar Saif and Rumi Chunara},
  title = {{Deep Landscape Features for Improving Vector-borne Disease Prediction}},
  booktitle = {Workshops @ CVPR},
  year = {2019},
  url = {https://openreview.net/forum?id=iyAWdqQLrz3},
}
Figure