Mapping canopy cover for municipal forestry monitoring: Using free Landsat imagery and machine learning

Abstract

Trees across the urban-rural continuum are recognized for their ecological importance and ecosystem services. Municipalities often utilize spatial canopy cover data for monitoring this resource. Monitoring frameworks typically rely on fine-scale maps derived from very high spatial resolution sensors, which are high quality but expensive and unwieldy for consistent wide-area monitoring. In this paper, we explore how free Landsat imagery, supported by very high-resolution imagery interpretation and/or digital hemispherical photographs, can be used to effectively map canopy cover at a scale appropriate for municipal monitoring. We compare linear models and random forest machine learning for predicting canopy cover across a landscape (general) and within specific land covers (specialized). We create 2018 canopy cover maps and track progress towards forestry objectives in a region of southern Ontario, Canada. Random forest models using all reference data perform best for general use (R2: 0.90, RMSE: 10.1 %), separating non-canopy vegetation (e.g., agricultural fields) from tree canopy. Specialized models are useful in forest land cover patches, where hemispherical photographs relate with Landsat at a moderate strength (R2: 0.67, RMSE: 2.73 %), and in residential areas, capturing the totality of canopy cover variation (R2: 0.85, RMSE: 5.66 %). Accuracy was assessed with standard cross-validation, which is useful given limited resources. However, following best practice, an independent reference sample was also leveraged to assess the best general model (R2: 0.86, RMSE: 11.4 %), indicating that cross-validation was slightly overoptimistic. Results show that Caledon, a rural-dominant municipality within the study area, is the greenest (34 % canopy cover). The two cities (Brampton and Mississauga) have 15.9 % and 17.5 % canopy cover. Residential canopy criteria indicate “Good” performance in Caledon, “Moderate” in Mississauga, and “Low” in Brampton based on our 2018 assessment. The methods described here can provide municipalities with a low-cost approach for tree canopy monitoring across complex landscapes. Data availabilityThe data used for this paper are available at: https://zenodo.org/records/12549244. The code created for this paper are available at: https://github.com/ZZMitch/PredictTreeCC_Landsat_1972to2020.