Abstract
Background Urban tree planting initiatives are a popular way to increase municipal tree presence. Initiatives on private land, including backyard tree planting programs, are essential because most available planting space across cities is on private property. Therefore, understanding the success of these programs, including long-term tree retention rates, is crucial for determining future urban forest characteristics and associated ecosystem services. However, few studies evaluate the outcome of backyard planting programs, primarily because of barriers like limited organizational resources and the inaccessibility of trees planted in backyards. To address this issue, our study examined the feasibility of using publicly available aerial imagery to assess long-term retention of trees planted through a backyard tree planting program in Toronto, Ontario.
Methods Using 20 years of leaf-off imagery and hand-drawn planting maps, a sample of 2,654 trees was assessed for feasibility of location digitization, presence-absence classification in 2022, and 5-year survivorship.
Results We successfully digitized 1,865 (70%) of these trees, but the remaining 30% could not be mapped due to insufficient location information. Of those digitized, we could confidently determine if 1,533 trees (82%) were present or absent in 2022. The status of the remaining 18% of trees was unclear, often due to image obstruction or quality. We were able to determine presence/absence 5 years after planting for 81% of trees in the subset examined.
Conclusions Ultimately, using aerial imagery could be a time- and cost-effective approach to long-term, ongoing urban tree monitoring, though challenges associated with image availability and quality should be considered.
Introduction
While cities are growing globally in both area and population (Angel et al. 2011; United Nations Department of Economic and Social Affairs 2019), they also face increasing pressures from climate change and urbanization. To mitigate these pressures, many cities are prioritizing the growth of their urban forests due to the ecosystem services they provide, such as carbon dioxide sequestration and local cooling (Moss et al. 2019; Steenberg et al. 2023). To achieve this growth and meet associated canopy cover or tree planting targets (Ordóñez and Duinker 2013; Ordóñez Barona et al. 2024), tree planting initiatives (TPIs) have emerged as a common approach, offered by municipalities, nonprofits, and community organizations (Greene et al. 2011; Nguyen et al. 2017; Elton et al. 2022; Healy et al. 2023). While many TPIs focus on planting trees on public property to maintain responsibility for stewardship practices crucial for tree survival (Roman et al. 2015; Doroski et al. 2020; Eisenman et al. 2021), those that target private property are as important because most of the remaining available planting area in cities is located on private land (City of Toronto 2018; Monteiro et al. 2020).
Backyard tree planting programs are one approach to increasing tree presence on private property. They often provide free or subsidized trees to homeowners (Roman et al. 2014; Ruseva et al. 2015; Morgan and Ries 2022) and offer guidance on tree planting and species selection (Greene et al. 2011; Nguyen et al. 2017). Such programs each typically plant hundreds of trees annually (Nguyen et al. 2017). Research into who participates in these programs has found that motivations for participation range from the aesthetics associated with trees to environmental benefits, such as providing habitat for wildlife (MacDonald et al. 2020; Geron et al. 2023). Studies have also suggested that these programs contribute to urban forest inequity (Donovan and Mills 2014) because trees tend to be planted in areas with higher educational attainment levels (Greene et al. 2011; Roman et al. 2014; Ruseva et al. 2015), though this is not true of all programs (Watkins et al. 2016; Nguyen et al. 2017).
While understanding program uptake and participation is important for addressing equity considerations and informing program development, understanding the long-term survival of trees planted through these programs is also a crucial measure of success. Trees need to reach a mature age and size to realize many of the ecosystem services they are planted to provide (Sawka et al. 2013; Lai et al. 2020; MacDonald et al. 2020). This is not only important for residents who plant trees, but also for the ability of TPIs to contribute to their own or municipal urban forestry goals and to ensure expected benefits for funding partnerships. However, due to limited organizational capacity to complete ongoing monitoring (Roman et al. 2013; Nguyen et al. 2017; Breger et al. 2019) and inaccessibility of backyard trees (Roman et al. 2014; Ko et al. 2015a; Nguyen et al. 2017), few studies explore the long-term survival of trees planted through backyard tree planting programs.
Research on the success of urban TPIs, including those focused on private trees, often relies on site visits to assess tree health, growth, and survival (Koeser et al. 2014; Roman et al. 2014; Ko et al. 2015a; Vogt et al. 2015; Elmes et al. 2018). While this allows for tree attributes to be accurately recorded, it is timeconsuming, with authors reporting multiple years’ worth of seasonal site assessments for samples often less than 500 trees (Nowak et al. 1990; Roman et al. 2014; Ko et al. 2015b; Bigelow et al. 2024). This indicates it may not be a feasible approach to ongoing, consistent monitoring of tree growth and retention, especially for organizations that do not have the internal capacity or academic partnerships to facilitate this monitoring, or where access to trees is dependent on resident cooperation.
A limited number of studies have employed remotely sensed imagery to assess backyard tree presence (e.g., Ossola and Hopton 2018; Pedley and Morgenroth 2025), and fewer have used aerial imagery to do so (Sawka et al. 2013; Ko et al. 2015a). However, these studies typically only assess survival at one point in time, focus on a relatively small sample of trees, or rely on older, lower-resolution imagery. Further, no identified studies explicitly address the challenges or best practices associated with this approach, particularly for ongoing longitudinal assessments. This limits the scalability and practical applicability of these methods for long-term monitoring of backyard tree planting programs or other urban TPIs. It also limits our understanding of how using imagery to assess tree retention may bias study results relating to mortality rates or other program outcomes like tree growth or resulting ecosystem services.
To address these limitations, this study explored the feasibility of using high-resolution aerial imagery to monitor the long-term retention of trees planted through a backyard tree planting program in Toronto, Ontario. Our research objectives were to: (1) develop a working procedure and classification system for assessing tree retention with aerial imagery; (2) identify strengths and weaknesses associated with this approach; and (3) provide recommendations for implementing these methods to organizations that may benefit from them. By focusing on a large-scale longitudinal retention assessment over 26 years, this study demonstrates the potential for remote sensing approaches to overcome common barriers to monitoring trees on private property and offers a scalable technique that can help inform the design and evaluation of TPIs.
Materials and Methods
Study Area
The City of Toronto is Canada’s largest municipality by population, with 2.8 million people and a population density of 4,428 people per square kilometer (Statistics Canada 2022). Toronto’s estimate of citywide average canopy cover is 28%, though this varies substantially by neighborhood (City of Toronto 2018). Approximately 33% of Toronto’s land area is residential (Toronto City Planning 2021), with an average canopy cover of 31% (City of Toronto 2018). Toronto is in the Mixedwood Plains Ecozone, characterized by cool winters, warm summers, and 720 to 1,000 mm of annual precipitation (Crins et al. 2009). While Toronto’s urban forest contains a mix of native and non-native species, the most abundant species are eastern white cedar (Thuja occidentalis), sugar maple (Acer saccharum) and Norway maple (A. platanoides) (City of Toronto 2018).
Planting Program Data
This study assessed the retention of trees planted by the nonprofit organization Local Enhancement and Appreciation of Forests (LEAF) through their Backyard Tree Planting (BYTP) program (LEAF 2011). This program aims to grow Toronto’s urban forest by providing native trees to homeowners at a subsidized price and has been offered in the City of Toronto since 1996. The BYTP is a full-service program which includes a consultation with a LEAF arborist to select an appropriate tree species and planting location according to residents’ yards and motivations for tree planting. Trees are planted in the agreed upon location by LEAF in the spring or fall following consultation.
Trees planted in single-family residential properties in the City of Toronto were assessed for retention in this study. Between 1996 and 2023, LEAF planted 6,920 trees through their BYTP program on Toronto residential properties. We selected a random sample of 2,000 trees for this analysis, proportionately stratified by 5-year planting intervals (e.g., 1996 to 2000) to ensure representation of trees of varying ages. Because LEAF often plants multiple trees per property at the same time (i.e., multiple trees in a single order), we included all additional trees from the same order as those in our original sample to allow for future assessment of how property-level characteristics affected tree retention.
For every order completed for LEAF’s BYTP program, program staff created a siting form. They consisted of a hand-drawn reference map of the yard and location where trees were planted, as well as contextual information about tree species, address of residence, and surrounding biophysical characteristics such as sunlight exposure, soil type, and housing structure. Each map included planting locations relative to objects in the backyard (e.g., building, fence, tree) which served as reference points. The distance from each tree to these objects was measured in feet or meters. These siting forms were digitized and provided by LEAF for all orders associated with the sample trees.
Spatial and Remote Sensing Data
Orthorectified aerial imagery accessed through the City of Toronto’s Open Data Portal was used to assess tree location and retention (City of Toronto 2024a). Imagery was available for 11 years between 2005 and 2022. Images were leaf-off, consistently taken between March and May, except in 2012 and 2016, which were taken between late April and June and were often leaf-on. Resolution from 2015 onwards was 8 cm but ranged from 5 cm to 20 cm in years prior. The projection of all pre-2020 imagery was NAD27 MTM, and from 2020 to 2022 was NAD83 CSRS MTM10. All planting locations were digitized to 2022 imagery, which had horizontal and vertical accuracies of 0.16 m and 0.24 m, respectively.
A point shapefile of all addresses in Toronto (City of Toronto 2024b), also accessed through the City of Toronto’s Open Data Portal, was used to locate the residences where trees were planted. All imagery and the Toronto address shapefile were accessed in April 2024.
Digitizing Planting Locations
The first step of the analysis was to digitize tree planting locations. The residences where trees were planted were identified using the address shapefile, LEAF’s siting forms, and Google Maps. Identified yards were visually compared to siting form maps to ensure they aligned, using the available imagery from the year closest to when the tree was planted.
Reference points and measurements from LEAF’s siting forms were then used in conjunction with 2022 aerial imagery to digitize planting locations (Figure 1). Earlier imagery (2005 to 2021) was used to confirm whether reference points visible in 2022 were correct. When original objects used as reference points were not visible in 2022, for example if they were removed or obscured, they were identified in earlier imagery and new measurements were created for reference points visible in 2022.
Example of the process used to digitize planting locations of trees, including original siting forms from Local Enhancement and Appreciation of Forests (LEAF) and imagery available through the City of Toronto (City of Toronto 2024a). (A) Example LEAF siting form, including hand-drawn map of planting location for a red maple and white pine tree planted in 1999. (B) Example of 4 years of available imagery used in this study, from earliest available (2005, 20-cm resolution) to most recent (2022, 8-cm resolution). (C) Example of final digitized locations, georeferenced to the 2022 imagery using the LEAF siting form and historic aerial imagery.
When a planting location had only one reference point, the tree was not digitized unless the location could be seen clearly in imagery for the year following planting and the existing measurement aligned with the visible location. When there were discrepancies between where siting form measurements placed trees and where it was clear a tree had been planted in available imagery, the location in the imagery was used if the discrepancy was less than 3 m. When discrepancies were greater than 3 m, the measurements from the siting form were used to digitize the planting location, and it was assumed that the tree was not visible on the imagery. Planting locations that could not be digitized were classified into reasons why, using a posteriori coding to identify limitations of this approach associated with digitizing locations from maps and reference points.
Two authors digitized and verified tree planting locations. Verification involved cross-checking the siting form against the digitized location. Any discrepancies were discussed until agreement on the correct location was reached. These planting locations were then used to assess tree retention across years of available imagery. Supplemental planting information from each siting form, such as soil type, sunlight exposure, and housing structure was recorded in the planting location shapefile for future analysis.
Retention Classification
Each tree with a successfully digitized planting location was then assessed for long-term retention by classifying whether it was still present based on the most recent (2022) imagery. Each tree was classified as either (1) Present; (2) Absent; or (3) Undetermined based on set criteria (Table 1; Figure 2). The terms present and absent over alive or dead trees, as well as retention rate over survival rate, were used because alive trees could not be differentiated from standing dead trees using leaf-off imagery. Trees coded as undetermined were further classified based on the reason the status was not clear using a posteriori codes to identify challenges associated with assessing long-term retention using aerial imagery.
Examples of trees classified as present, absent, and undetermined in 2022, including the 2022 imagery (City of Toronto 2024a) using the criteria outlined in Table 1. (A) Example of a tree classified as present (hackberry tree planted in 1996). (B) Example of 2 trees classified as absent (white pine and serviceberry trees planted in 2012). (C) Example of a tree classified as undetermined (serviceberry tree planted in 2017).
Criteria for classifying trees as present, absent, or undetermined for a given year of assessment using high-resolution historic orthorectified aerial imagery.
All trees were independently classified as present, absent or undetermined by two of the authors. Inter-rater reliability for classifications was calculated using Cohen’s Kappa Statistic (k)(Equation 1) (McHugh 2012):
Where k is the Cohen Kappa Statistic, po is the relative observed agreement among reviewers, and pe is the hypothetical probability that reviewers may have agreed on a classification by chance. This value can range between -1 and 1, where values between 0.61 and 0.80 indicate substantial agreement, and those above 0.81 indicate almost perfect agreement (McHugh 2012). After reliability was calculated, classifications were compared, and discrepancies were discussed until consensus was reached to provide the final classification of trees displayed in the results.
To help determine how retention results identified using this method compare to broader literature on urban tree mortality, we also assessed 5-year retention rates. This was done for all sample trees where at least 10-cm resolution imagery was available exactly 5 years after the trees were planted, which included trees planted in 2004, 2006 to 2007, 2010 to 2013, and 2015 to 2017. Again, trees were classified as present, absent, or undetermined 5 years after their planting using all years of available imagery. This was done using the criteria in Table 1, with one additional way for trees to be classified as present: if it was unclear whether a tree was present in the imagery exactly 5 years after it was planted, but could be clearly seen in later imagery, it was still classified as present 5 years after planting. Annual removal rate was calculated using Equation 2 (Nowak et al. 2004; Lawrence et al. 2012):
Where N1 is the number of trees remaining after 5 years, N0 is the number of trees planted, and t is the number of years between assessments.
Methodological Success by Planting Characteristics
Whether a tree could be classified as present, absent, or undetermined was assessed against tree and propertylevel characteristics. This was done to provide further context into the application of this methodology across species and biophysical environments and to identify how this method may bias future analyses run with the resulting dataset. A binary variable was created representing whether trees’ presence/absence classification could be determined in 2022 (1 = yes) or if the status was undetermined (0).
Five variables were selected to model classification success based on associated characteristics anecdotally observed during our analysis and that have been associated with tree planting or mortality patterns in existing research (Table 2). Years since planting, coniferous vs. deciduous species, and number of trees planted per order were derived from the siting form. Residential property size was derived from the City of Toronto’s Open Data Portal Property Boundary dataset (City of Toronto 2024c). Property-level canopy cover was sourced from the City of Toronto’s Open Data Portal Forest and Land Cover dataset (City of Toronto 2022) and was calculated as the percent of a property classified as tree cover.
Independent variables used to model success in classifying the presence and absence of backyard trees in 2022, including the observed and theoretical rationale for inclusion, associated references, and summary statistics for our sample.
Analysis to identify factors significantly associated with the success of classifying trees was done using logistic regression in R v.4.4.0 (R Core Team, Vienna, Austria) using the glm function in the stats package. The model was validated using the Akaike information criterion (AIC)(summary function in base R package), the area under the curve (AUC) value (roc function in the pROC package)(Robin et al. 2011), Hosmer-Lemeshow test (hltest function in glmtoolbox package)(Vanegas et al. 2024), deviance model residuals (summary function in base R package), and was assessed for multicollinearity using a variance inflation factor (VIF) threshold of 3 (vif function in cars package)(Fox and Weisberg 2019). During model validity assessments, the variable representing years since planting was log-transformed to improve model fit.
Results
Overview
A total of 2,654 sample trees were included in this analysis: 2,000 from the original sample and 654 that were in the same order as original sample trees. This sample encompassed 47 unique species, with eastern white cedar (Thuja occidentalis, n = 318), green ash (Fraxinus pennsylvanica, n = 196), and red maple (Acer rubrum, n = 170) being the most common. Sampled trees spanned 25 unique planting years from 1996 to 2023 and were distributed across the city (Figure 3). It took approximately 10 minutes per reviewer to digitize and classify each tree.
Residential addresses of the 2,654 sample trees planted through Local Enhancement and Appreciation of Forests’ (LEAF) Backyard Tree Planting program from 1996 to 2023.
Digitized Planting Locations
Of the 2,654 sample trees, we were able to digitize the planting locations for 70% (Table 3) across 1,297 orders. There were 6 reasons that planting locations were not digitized, with most associated with limitations of the siting forms.
Breakdown of sample trees based on if their planting location was digitized (Yes) or was not digitized (No) and the codes developed for reasons that locations could not be digitized.
The 2 most common reasons account for 62% of unmapped trees. For 295 of these trees, the associated siting form had no (legible) measurements to reference points or contained only one measurement and the planting location was not visible in any available imagery. For 197 of these trees, 2 measurements were present on the siting form map, but the objects used as references were not identifiable in any available imagery. Sometimes these were larger reference points that had been removed or replaced, such as sheds or decks, predominantly affecting trees planted before 2005. More commonly, smaller or less permanent objects used as reference points (e.g., rocks, birdbaths, existing trees) caused this issue.
Retention Classification
Of the 1,865 trees whose planting locations were digitized, 44% were classified as present and 38% as absent in 2022 (Table 4). However, presence or absence for 18% of the trees could not be determined. Percent agreement for classifying presence/absence status in 2022 was 82%, and reliability calculated using Cohen’s Kappa Statistic was k = 0.72, indicating substantial agreement.
Breakdown of presence/absence status in 2022 of the 1,865 sample trees and classifications for why this status could not be determined for some trees.
While all 3 reasons that presence/absence could not be determined were related to the aerial imagery used, the most common classification, affecting 228 trees, was that the planting location was obstructed in 2022 (Table 4). Locations were often obstructed by large canopies of older trees or shadows from buildings and other structures that made the imagery very dark. In addition to image obstruction, there were also issues associated with image quality that made it difficult to determine tree presence/absence. This occurred frequently in areas with other vegetation or ground cover. Both issues predominantly affected trees that were smaller in stature or were more recently planted.
Regarding establishment period survival, 748 (81%) of trees planted in the included years could be classified as present or absent 5 years after planting based on image availability. Of those classified as present or absent, 620 (83%) were present 5 years after planting, corresponding to an annual removal rate of 3.7%.
Classification Results by Tree and Property-Level Characteristics
Three biophysical planting characteristics were significantly associated with the likelihood that a tree’s presence/absence status could be determined in 2022 (Table 5). For every e-fold increase (approximately 2.72 times) in the number of years since a tree was planted, the odds its status could be determined increased by 3.78 (P < 0.001). For example, across our sample with trees planted between 1996 and 2021, this means that the odds of determining a tree’s status were approximately 76 times higher for a tree planted in 1996 than in 2021. For every additional tree planted in the same order, the odds a tree’s status could be determined increased by 1.12 (P = 0.012). For every 1% increase in property-level canopy cover, the odds a tree’s status could be determined decreased by 0.04 (P < 0.001). For every square meter increase in the size of the property where trees were planted, the odds the tree’s status could be determined increased by 0.001 (P < 0.001). Whether a species was coniferous or deciduous did not significantly affect the likelihood that status could be determined (P = 0.850).
Odds ratio estimates from logistic regression for biophysical planting characteristics associated with the likelihood that the presence/absence of a tree could be determined in the year 2022. Odds ratio estimates greater than one indicate an increased likelihood status could be determined.
Discussion
This study aimed to assess the strengths, challenges, and ultimate feasibility of using aerial imagery to monitor retention of trees planted through urban backyard tree planting programs. Using 2,654 trees planted through LEAF’s BYTP program, we found that it can be a cost-effective approach to monitoring, but that it is not without challenges. We were able to digitize the location of 70% of our sample trees and classify 82% of those as present or absent in the year 2022. Success with both location digitizing and classification depended on the quality of program documentation, tree characteristics, and the biophysical characteristics of the property the trees were planted in. This discussion will outline the observed strengths and weaknesses, as well as recommendations for applying this approach to similar programs.
Strengths
Compared to existing studies on urban tree survival, we found that using aerial imagery to monitor retention appears both cost- and time-effective. Similar studies that have used site visits to monitor tree survival have reported 2 to 4 summer months of field work for monitoring between 297 and 4,059 trees (Jack-Scott 2012; Roman et al. 2014; Ko et al. 2015a, 2015b; Widney et al. 2016; Breger et al. 2019; Bigelow et al. 2024). While most studies do not further discuss time or capacity restraints associated with monitoring, some outline needing to involve and train students or volunteers to complete sampling (Widney et al. 2016; Breger et al. 2019; Hauer et al. 2020). Another study that assessed the time required to collect similar monitoring data from i-Tree plots found that assessments took between 16 and 24 minutes per tree, depending on plot size (Nowak et al. 2008). However, while their estimate included setup, surrounding area assessments (e.g., surrounding canopy cover), and individual tree measurements, it excluded travel time. While our approach required an initial time investment to digitize tree locations, approximately 7 minutes per tree, it is likely still more efficient than field visits distributed across a city, does not need to be repeated for follow-up monitoring, and would not apply to programs that record tree locations as coordinates at planting. Identifying tree presence/absence status in 2022 took approximately 3 minutes per tree after locations had been digitized and would likely be faster for subsequent years of monitoring. Ultimately, this method appears more scalable and manageable in terms of time efficiency and staff or volunteer capacity.
In addition to being a potentially more efficient method of analysis, this approach yielded results that align with existing research on urban tree mortality rates assessed using site visits. Of trees successfully classified as present or absent 5 years after planting, 83% were retained with an annual removal rate of 3.7%. This is similar to but slightly lower than annual mortality rates observed using site visits in residential trees in Sacramento, California, USA, at 6.6% per year (Roman et al. 2014). More broadly, studies that have assessed 5-year survival rates have found median mortality rates between 6.6% and 7% per year (Hilbert et al. 2019). While our lower observed rate may be due to biases associated with the inability to successfully classify all trees for 5-year retention, it may also be due to the design of LEAF’s BYTP program. It prioritizes selecting planting locations that minimize existing and potential future conflicts for trees, planting species that align with biophysical conditions of planting locations, considering homeowner motivations for tree planting in both site and species selection, and providing tree care resources to program participants.
In addition to serving as a stand-alone method for long-term monitoring, this approach could supplement periodic site visits to a subsample of trees to provide a more comprehensive understanding of tree removal dynamics. Unlike site visits, which typically offer detailed but infrequent snapshots of tree survival, annual analysis using aerial imagery could reveal nuanced trends in tree retention, removal, and replacements that may otherwise go unnoticed. It could provide smaller, more accurate timeframes for when trees were removed while providing additional context about environmental factors or management practices that contribute to these removals (e.g., land use change), as has been demonstrated in existing research (Ko et al. 2015a). This combined approach would also increase the level of certainty associated with classifications made using imagery alone and could help better contextualize the results of this study.
Another benefit of this approach is that it is accessible across cities, as many, including Philadelphia, Pennsylvania, USA (City of Philadelphia [date unknown]), Los Angeles, California, USA (County of Los Angeles [date unknown]), and New York City, New York, USA (City of New York 2024), offer publicly accessible imagery with resolutions comparable to those used in this study. Given that these cities have TPIs that are the focus of several studies on urban tree mortality (Morani et al. 2011; McPherson 2014; Roman et al. 2015; Bigelow et al. 2024), this approach would allow programs or researchers to complete more frequent and efficient monitoring without the need for annual field visits. For cities without open-source imagery, resources such as Google Earth (Google [date unknown]), EarthExplorer (US Geological Survey [date unknown]), or Planet (Planet Labs [date unknown]) can be used to access high-resolution imagery globally, though often for a cost.
Challenges
While there are strengths associated with this methodology, there are also many limitations. As identified through descriptive coding, issues associated with image obstruction and quality resulted in the inability to assess tree presence/absence for hundreds of trees. Existing studies that have used aerial imagery to assess tree retention have used images with resolutions ranging from 15 cm to 1 m to confidently classify survival status of trees between 5 and 20 years old (Ko et al. 2015a, 2015b). The difference in success classifying trees at different image resolutions could be due to differences in canopy cover and other vegetation presences between Toronto, Ontario, and Sacramento, California, USA, the focus of Ko et al.’s (2015a, 2015b) studies. While those studies were supplemented with 2 field visits, which also could have helped increase confidence in classification between visit years, we found that even with image resolution of 8 cm, it can be difficult to assess retention, particularly for trees under 5 years old in yards that include other vegetation and building shadows.
In addition to image resolution, historic image availability and obstructions associated with the time of day or year images were taken also introduced limitations in determining tree status. Because our sample included trees planted in 1996 onwards, and the earliest year imagery was available was 2005, there was a 9-year period where we could not assess tree removal dynamics. This means that had trees been replaced in this timeframe, we would not have been able to tell and would likely have classified that tree as present. This limitation is not unique to aerial imagery, however, and has been noted as a consideration when relying only on site visits as well (Ko et al. 2015a). Similarly, if trees were moved from the planting location specified in their siting form, as was suspected for 13 trees, this may also have gone unnoticed and inflated the number of trees classified as absent.
Further, certain biophysical characteristics of trees or the area surrounding planting locations limited our ability to classify trees. In addition to younger trees, it was more difficult to identify the status of trees planted in areas with higher existing canopy cover, as this canopy often obstructed planting locations. This may be an increasing issue as cities continue to plant trees to reach canopy cover targets (Ordóñez and Duinker 2013; Ordóñez Barona et al. 2024). However, this finding also suggests that this method could be particularly useful for monitoring program success in areas with disproportionately low canopy cover, which are the focus of several urban TPIs (Watkins et al. 2016; Nguyen et al. 2017). We also found that trees were harder to classify as property size decreased, likely because the large properties in our sample were largely open space, free of shadows from surrounding buildings or other structures. However, the low odds ratio estimate suggests this effect is most pronounced on very large lots and may have limited influence on most properties, especially given the wide range of property sizes in our sample. Further, when fewer trees were planted in an order, it also became less likely the status of a tree could be determined. This may have been because residents planting many trees (e.g., > 10) often had no other vegetation in their yards at the time of planting and planted few trees after those planted by LEAF. This made it easier to consistently identify trees across years of imagery with very little uncertainty about if trees were the correct species, age, or at the correct planting location. This also supports the applicability of this methodology to residential areas with low initial canopy cover, as well as its use to monitor commercial or institutional planting that typically plant several trees on a property at once.
Another limitation associated with relying on aerial imagery is the inability to assess other key metrics of program success, such as tree growth or health. These factors are important determinants of ecosystem service delivery (Nowak and Crane 2002; Boyd et al. 2013) and are often associated with tree mortality (Das et al. 2007; Kirkpatrick et al. 2013; Ko et al. 2015b; Conway 2016). Encompassing multispectral data such as LiDAR could help address this limitation, providing more insight into annual growth rates and tree health. Future research should explore these opportunities, particularly as high-resolution multispectral datasets become increasingly available.
In addition to limitations with aerial imagery, human error is also associated with this approach. Those responsible for classifying presence/absence of trees should be familiar with tree morphology, able to distinguish between tree species, and able to recognize appropriate characteristics of the trees being assessed. Lacking this understanding may result in an inflation of presence or absence classifications depending on reviewer knowledge. The subjective element of classification is also apparent in our level of inter-rater reliability of 0.72, even though it aligns with cross-disciplinary standards for acceptable agreement rates (McHugh 2012).
Recommendations for Similar Programs
This study has several implications for urban TPIs that are interested in best practices associated with long-term monitoring of trees planted through their programs. In cases where tree locations are recorded through hand-drawn maps, at least 2 reference points should be used to ensure planting locations can be digitized. All reference points should be stable locations such as fences along property boundaries or houses. Existing trees should not be used, nor should small features or those that are likely to be moved or redesigned such as trampolines or decks. Further, there must be an emphasis on correct, complete, and consistent records across program documents.
To help ensure accurate and unbiased results, it is important to have 2 people independently classify the status of each tree and to have access to multiple years’ worth of imagery. Interpreting aerial imagery is an inherently subjective process (Lillesand et al. 2015). While inter-rater reliability in this study was considered sufficient, there were still over 100 trees that reviewers initially classified differently. This suggests that factors like reviewers’ past experiences or confidence levels can bias the distribution of classification results, potentially impacting key TPI outcome estimates. Having at least 2 people classify each tree can help reduce these biases, especially if reviewers can discuss and rectify differences. This subjectivity has also been observed in tree risk assessment research that shows even experienced arborists draw different conclusions about the same trees without added complications from aerial imagery (Koeser and Smiley 2017).
Having multiple years’ worth of available imagery, ideally ranging from when the trees were planted to the year of assessment, was important for reducing the number of trees classified as undetermined. We noticed that image quality and obstruction levels varied for the same properties across years, resulting from differences in resolution, the time of day and year images were taken, and differences in backyard structure. This variation meant tree presence or absence was clearer in some years, which could be used to help inform classifications in 2022. For example, if it was unclear whether a tree was present in 2022, but it was very clear in earlier imagery that the tree had been removed, it could confidently be classified as absent in our analysis. Further, it was sometimes difficult to tell whether trees visible in 2022 were the correct ones from a single year of imagery. Clearly seeing a tree when it was young and identifying it in subsequent imagery to 2022 helped reduce these uncertainties and false classifications, for example by allowing us to identify when trees may have been replaced.
Conclusions
This study demonstrated that high-resolution aerial imagery can be used to monitor long-term retention of trees planted through urban backyard tree planting programs. By developing and testing a scalable methodology for long-term monitoring, identifying associated strengths and challenges, and addressing program-level factors that can help increase the success of this method, this research offers a practical tool for monitoring urban tree retention. While limitations associated with aerial imagery obstruction and quality exist, integrating this approach with field surveys or multispectral data can help enhance monitoring efforts and understanding of factors that affect program outcomes.
Future studies should focus on applying these methods to other types of urban TPIs to better understand how they can be used to support urban greening initiatives. Datasets created through these types of studies could also be used to train deep learning models that could automate tree identification processes to further reduce the time associated with and increase scalability of this approach to monitoring. Similarly, these datasets can also be used to help identify different biophysical factors associated with the retention and removal of trees planted through urban TPIs.
Conflicts of Interest
Tenley Conway reports serving in a voluntary role with LEAF (Local Enhancement and Appreciation of Forests). Janet McKay reports serving as Executive Director of LEAF. The remaining authors reported no conflicts of interest.
Acknowledgements
Thank you to Erin McDonald and LEAF volunteers for digitizing the siting forms. Funding was provided by the Social Sciences and Humanities Research Council of Canada (project no. 435-2017-0093).
- © 2025 International Society of Arboriculture
Literature Cited
Jump to section
Related Articles
Cited By...
- No citing articles found.