Submission ID 118168
| Issue/Objective | This study addresses limitations in traditional methods that associate green space with health using the Normalized Difference Vegetation Index (NDVI), which cannot distinguish vegetation types (e.g., tree canopy versus grass). By integrating high-resolution multispectral satellite imagery with Light Detection and Ranging (LiDAR) data, this research aims to improve land use land cover (LULC) classification in Mississauga, Ontario. The primary objective is to develop a random forest supervised classification model that accurately delineates various green space types (e.g., deciduous trees, coniferous trees, shrubs, grass) and assess their association with diabetes prevalence. We hypothesize that quantifying the area of each green space type per census tract will predict diabetes prevalence more effectively than NDVI alone, and that stratifying by sex and age will reveal nuanced health patterns. |
|---|---|
| Methodology/Approach | This study uses multispectral satellite imagery (3-metre spatial resolution) and LiDAR data. The LiDAR data will generate normalized Digital Surface Models (nDSM) to capture object heights. Eight vegetation indices will be calculated from the satellite imagery. These indices, along with the nDSM and the satellite imagery bands, will create eight band combinations to develop object-based random forest classification models. Training samples will be manually created using reference imagery. Confusion matrices will identify the most accurate band combination for final LULC classification. Regression models will examine the association between the quantified green spaces and diabetes prevalence, controlling for socio-demographic and economic factors, with comparisons made against NDVI-based models. |
| Results | Preliminary results show that incorporating LiDAR data enhances classification accuracy by 6-12%, and the model using all indices, the nDSM, and the bands displays the highest overall accuracy (88%). It is anticipated that quantifying green space types will reveal a stronger, statistically significant association with diabetes prevalence (higher R2 values) compared to NDVI alone, while also revealing stratified differences by sex and age. |
| Discussion/Conclusion | Integrating multispectral satellite imagery and LiDAR data improves the classification of green spaces, enabling a more precise assessment of their health impacts. This study can inform urban planning and public health policies by emphasizing detailed LULC classification in understanding environmental determinants of diabetes. The methodology offers a scalable framework for other urban regions. |
| Presenters and affiliations | Scarlett Rakowska University of Toronto Matthew Adams University of Toronto |
