Remote Sensing in Urban Forestry: Recent Applications and Future Directions

Increasing recognition of the importance of urban forest ecosystem services calls for the sustainable management of urban forests, which requires timely and accurate information on the status, trends and interactions between socioeconomic and ecological processes pertaining to urban forests. In this regard, remote sensing, especially with its recent advances in sensors and data processing methods, has emerged as a premier and useful observational and analytical tool. This study summarises recent remote sensing applications in urban forestry from the perspective of three distinctive themes: multi-source, multi-temporal and multi-scale inputs. It reviews how different sources of remotely sensed data offer a fast, replicable and scalable way to quantify urban forest dynamics at varying spatiotemporal scales on a case-by-case basis. Combined optical imagery and LiDAR data results as the most promising among multi-source inputs; in addition, future efforts should focus on enhancing data processing efficiency. For long-term multi-temporal inputs, in the event satellite imagery is the only available data source, future work should improve haze-/cloud-removal techniques for enhancing image quality. Current attention given to multi-scale inputs remains limited; hence, future studies should be more aware of scale effects and cautiously draw conclusions.


Introduction
The world is experiencing accelerated urbanisation and growth of cities [1,2], which have dramatically changed the urban landscape [3][4][5]. Though frequently escaping our notice, a mosaic of fragmented habitats exists, ranging from remnant natural, modified natural, semi-natural, emulated natural or entirely human-created, where diverse trees and other vegetation thrive [6,7]. Collectively, all woods and trees within a city boundary are defined as an urban forest, including individual trees, forest stands, and their associated biotic and abiotic components [8][9][10][11]. Complex forest patches are a critical component of urban ecosystems. They provide a full spectrum of ecosystem services, referring to the benefits that human populations obtain directly or indirectly [7,12,13] from natural ecosystems [14,15], which serve as the basis of economic development and social wellbeing [16][17][18]. Some well-recognised and well-studied examples of urban forest ecosystem services include alleviating summer heat through evaporation and shading [19,20], reducing stormwater runoff by intercepting and absorbing water and improving infiltration [21,22], enhancing air quality by removing gaseous pollutants and particulate matter [23], reducing greenhouse gas emissions by storing and sequestering a considerable amount of carbon [24][25][26][27]; addressing food security by supplying food, medicine and materials [28], pollination [29][30][31], outdoor recreation opportunities [1,32], as well as serving as important places for aesthetics and spirituality [33,34].

Remote Sensing Applications in Urban Forests
To summarise recent applications of remote sensing in the field of urban forestry, we focused our search on research articles published in peer-reviewed journals, in English, within the time span between January 2013 and March 2019. Multiple databases, including ISI Web of Science, Elsevier's Scopus, Google Scholar and ScienceDirect, were browsed using the terms "remote sensing OR LiDAR" AND "urban forest OR tree OR plant OR vegetation OR greenspace". More than 200 studies were found to satisfy the aforementioned conditions. These studies were further filtered based on the following criteria: (1) The research must focus exclusively on urban forests or greenspaces. Studies involving natural forests are excluded.
(2) The research must use remotely sensed data obtained from airborne/spaceborne imagery or LiDAR. Studies involving data based on other techniques, such as pure photogrammetry or terrestrial-based methods, are excluded. (3) The research must involve at least one of the three themes: multi-source, multi-temporal, and multi-scale inputs. Studies unrelated to any of these themes are excluded.
The final number of studies found totalled 63; these were classified in terms of 10 different applications and three distinctive themes (Table 1). A detailed description of each study is reported in Supplementary Materials. The articles cover remote sensing applications in diverse areas, including measurement of urban forest attributes (e.g., biomass, species composition, canopy cover and spatial configuration) and quantification of ecosystem services (e.g., carbon storage, water retention, air quality improvement, and urban heat island mitigation). In the following sections, the articles are reviewed and discussed in terms of the themes multi-source, multi-temporal and multi-scale input; their strengths and limitations are summarised in Table 2.

Theme 1: Multi-Source Input
A shift from the use of single-source remote sensing data (11 of 63) to multi-source remote sensing data (52 of 63) in urban forest studies has been observed thanks to the emergence of diverse airborne and spaceborne platforms and the rapid development of remote sensing techniques. Examples of these are the emerging very high spatial resolution (VHSR) and high spatial resolution (HSR) sensors, hyperspectral sensors and high point density LiDAR technology. Table 3 lists the remote sensing products/platforms used for urban forest studies. Two examples, as shown in Figure 1, are satellite imagery (WorldView-2 NDVI) and LiDAR point-cloud data (canopy height model) applied to urban forests at both macro-and micro-scale. Quantum leaps in computing power have enabled researchers to fuse, process and model data from various sources at an unprecedented level of spatial and temporal resolution.

Multiple Sources of Satellite Imagery
The combination of multiple sources of satellite imagery offers new potential in ecological research. In fact, the majority of studies addressing the link between urban greenspace and urban heat island rely on satellite images from multiple sources (Table 1). While HSR optical images, such as SPOT and IKONOS, are able to classify vegetation coverage and other land uses at a fine level, these optical sensors do not provide thermal infrared information. Indeed, land surface temperature has been derived from other images with thermal bands, such as Landsat and Terra/MODIS sensors. As satellite images from various sources complement one another, scientists overcome the technical challenges associated with a limited spatial, spectral, temporal, and radiometric resolution by creatively combining various satellite sources. Moreover, the increasing availability of free data is an incentive to utilise multiple sources of satellite imagery. Although there are certain advantages of fee-based imagery (e.g., very high spatial, temporal and spectral resolution) that cannot be replaced by open-access data, these data sharing platforms allow researchers to download a variety of satellite images from high-quality sources. Two such examples are the online platform operated by the United States Geological Survey (USGS) (https://earthexplorer.usgs.gov/) and the Sentinel online platform operated by the European Space Agency (ESA) (https://sentinel.esa.int/web/sentinel/home). Other ready-made global/regional land cover and land use products, such as CORINE and Geo-Wiki, can be directly used for analysis. Nonetheless, research should be the foremost consideration when it comes to the selection of satellite imagery.
For example, Ren et al. [51] examined the relationship between urban park features and park cool island intensity in Changchun, China using multi-source satellite imagery. The urban land surface temperature was derived from Landsat-5 Thematic Mapper (TM) thermal band using the mono-window method proposed by Qin et al. [52]. SPOT images were used to extract urban park spatial features like size and shape.
Kong et al. [53] used two sources of satellite imagery, IKONOS and Landsat-5 (TM) images, to explore the impacts of spatial greenspace pattern on the urban heat island effect in a large metropolitan area in eastern China. The HSR IKONOS images were used to classify land use into six types. Band 6 of Landsat-5 (TM) imagery was able to provide thermal infrared data that could be used to estimate land surface temperature. Aligned with the spatial resolution of the land use map, the land surface temperature map was resampled at 5 m × 5 m. The land use map was employed to calculate landscape indices, which were used to estimate the spatial pattern of greenspace and to analyse its relationship with land surface temperature. It was found that the greenspace pattern significantly influenced the urban cooling effect. Greater urban cooling effects were observed in areas with a high percentage of greenspace cover and level of greenspace fragmentation.
In addition to quantifying the urban cooling effect of urban greenspaces, multi-source remote sensing has also been recently applied to address other research objectives. Modugno et al. [54] mapped forest fires in the wildland-urban interface areas across Europe based on the harmonised land use and land cover change map CORINE 2006. The CORINE land cover product derived from multi-source satellite imagery is a commonly used dataset for European continental scale studies [55,56] and has been updated every six years since 2000. The CORINE 2006 dataset selected in this study is based on SPOT-4/-5 and IRS P6 satellite imagery. The forest fire information was extracted from the Burned Surfaces Product, which was derived from the European Forest Fire Information System using MODIS images. It revealed all forest fires burning 40 ha of surface area. Instantly accessible harmonised datasets (e.g., CORINE land cover product and burned surfaces product) based on multi-source satellite imagery have enabled researchers to evaluate attributes of urban forests with high efficiency.
In general, the application of the multi-source approach could be limited by two technical barriers. First, weather conditions (e.g., haze and cloud coverage) and different image dates; both pose major challenges and there is no guarantee that all images obtained within a certain time period are of high quality for classification purposes. For example, in their 20-year temporal analysis, Di Leo et al. [57] discarded poor-quality images affected by cloudiness or haze. The feasibility of using multiple sources of satellite imagery should be a primary consideration, especially for time-sensitive research. Second, satellite images are rarely available at both HSR and spectral resolution. Hyperion EO-1 imagery is one of the few hyperspectral satellite images with 242 bands, yet it only has a spatial resolution of 30 m. Till now, few tree-level studies (e.g., identifying individual tree species and tree crowns) have relied solely on multiple source applications of satellite images.

Satellite Imagery and Airborne LiDAR
Due to the strengths in providing spectral details and 3D point clouds, VHSR/HSR satellite imagery and airborne LiDAR data were suited for the estimation of urban forest structure, attributes and carbon storage. The integration of optical satellite imagery and LiDAR data allows us to identify the urban forest structure more accurately in a complex urban environment. HSR multispectral imagery is able to produce precise vegetation coverage, upon which LiDAR point-cloud data are based and can be further used to estimate tree-level structure.
For example, Raciti et al. [58] mapped carbon storage of urban trees in Boston City using LiDAR data coupled with VHSR satellite images. QuickBird imagery was first employed to derive the Normalised Difference Vegetation Index (NDVI) values for extracting the vegetation cover. Subsequently, 3D airborne LiDAR point-cloud data were used to delineate individual tree crowns within the vegetation coverage. LiDAR-based tree height data were further used to estimate biomass and carbon storage. The results showed that previous work using single-source remotely sensed data tended to overestimate carbon storage in Boston. The QuickBird-LiDAR approach presented in this study displayed high accuracy in estimating stereo features of urban trees, like volume and biomass.
Similarly, Schreyer et al. [59] developed a more specific method to combine LiDAR data and VHSR satellite images to evaluate the carbon storage of urban trees in Berlin, Germany. Airborne LiDAR was first employed to generate a digital surface model (DSM) and a digital terrain model (DTM).
After a vegetation mask was created based on QuickBird imagery, the DSM and DTM were used to derive a canopy height model (CHM). Using a two-step approach [60,61], dendrometric parameters including tree location, number, height, crown width and diameter at breast height were generated from the CHM. These parameters were then used to calculate carbon storage and distribution among urban trees.
Parmehr et al. [62] used a novel data fusion strategy to identify urban tree canopy cover with WorldView-2 imagery and airborne LiDAR data. The investigators adopted an automatic co-registration method, which was proposed by Parmehr et al. [63]. The co-registration of satellite imagery to LiDAR data is a process known as point-cloud colourisation. It attaches spectral information to the LiDAR point cloud so that each colourised LiDAR point has additional RGB and NDVI values. The NDVI values can improve the accuracy of identifying LiDAR tree points during the segmentation of trees from other urban objects. Data fusion of LiDAR point-cloud data and spectral imagery has great potential in increasing the accuracy of estimating tree structure in complex urban environments with diversified tree canopy shapes and terrains.
While most recent studies agree that the combination of satellite imagery and LiDAR data can lead to higher estimation accuracy, a publication by Singh et al. [64] challenges this consensus. The authors found that with the help of a random forest classification method, LiDAR-derived topography metrics could detect understory plant invasion in urban forests with the highest accuracy, exceeding IKONOS data by 17.5% and combined LiDAR and IKONOS data by 5.3%. In this case, they inferred that the temporal lag that took place between LiDAR and IKONOS data collection might have affected the accuracy of identifying the fast-growing invasive species, Ligustrum sinense. The off-nadir effect of tall trees [65] would also hamper HSR satellite imagery in detecting understory features in a heterogeneous urban environment.
Although some researchers argue that airborne LiDAR alone is capable of producing an accurate tree-level analysis, the abundance of spectral details provided by satellite imagery should be corroborated; managed properly it can significantly enhance estimation accuracy as opposed to using LiDAR data alone. Point-cloud colourisation is a promising technique for remote sensing data fusion. By combining every LiDAR point with spectral information extracted from satellite imagery, the colourised 3D point clouds with NDVI can be employed not only for segmenting individual trees but also for distinguishing trees from their surrounding urban environments. Nonetheless, it should be noted that the estimation accuracy of combining satellite imagery and LiDAR data would be reduced if the fetch interval between the two data sources is relatively long, or if the research objects are understory urban forests.

Satellite Imagery and Aerial Imagery
In practice, the combination of satellite imagery and aerial imagery has seldom been adopted, given that aerial imagery can replace satellite imagery in terms of optical capability. Only when extra data, such as infrared and thermal infrared data, are needed would such a combination become necessary. The majority of recent publications utilising both satellite and aerial imagery (three of five) have studied the mitigation effect of urban forests on the urban heat island. For example, Sung [66] estimated how the woodlands' tree protection policy adopted by The Woodlands Township, Texas, USA, affected the urban heat island at the neighbourhood level. The author compared the land surface temperature of two neighbourhoods in the woodlands and five control neighbourhoods in other urban areas. The land surface temperature was obtained from the thermal bands of Landsat-5 (TM) imagery. Unlike most urban cooling studies that used HSR satellite imagery to extract land use information, this study exploited the US National Agricultural Imagery Program's (NAIP) aerial imagery to classify land cover at a fine scale. An object-based method was coupled with a support vector machine (SVM) classifier to identify trees. When dealing with high-resolution imagery, object-based methods have been reported to be superior to pixel-based methods for their ability to reduce noise and increase accuracy [67,68]. Aerial imagery is superior to satellite imagery in terms of concurrent high spatial and spectral resolution. However, prohibitive flight costs limit its use to small/moderate areas. In the US, NAIP has been offering aerial imagery with continental coverage during the agricultural growing seasons to the public since 2003. NAIP imagery with four spectral bands is acquired at a 1-m ground sample distance with horizontal accuracy that matches within 6 m of ground control points, which can be used during image inspection and processing. A number of recent urban forest studies have benefitted from free NAIP images.

Airborne LiDAR and Aerial Imagery
The integration of airborne LiDAR and aerial imagery has recently gained popularity in tree-level urban forest inventories. It has great potential not only for enhancing estimation accuracy but also for increasing the number of identifiable tree species. However, the limited coverage and expensive flight costs of LiDAR hamper its use at a large scale. Alonzo et al. [69], for example, found that by using fused HSR hyperspectral AVIRIS imagery with high point density LiDAR data, 29 common urban tree species at crown level could be clearly identified. This was a substantial improvement compared to 15 species identified using AVIRIS imagery alone, as reported in an earlier study [70]. Even though the addition of LiDAR data could increase the overall accuracy by only 4.2%, it was able to enhance the classification accuracy of certain species, especially small crowned species, by more than 10%. Later, Alonzo et al. [43] found that fused airborne LiDAR and aerial imagery could outperform field-plot data when measuring the uneven distribution of canopies across the city. Compared to field-plot data, the consistency of airborne fused data is very suitable for measuring hard-to-access urban areas or for large-scale measurements that require replication. These studies confirm the merits of using hyperspectral imagery and LiDAR data fusion in mapping crown-level tree species in a highly heterogeneous urban environment.
Zhang et al. [71] designed a framework to integrate airborne LiDAR data and aerial images for a tree-level urban forest inventory. Above-ground points were separated from ground points in the LiDAR data. The authors innovatively used NDVI values derived from hyperspectral images to separate vegetation points from non-vegetation points. The ground points and vegetation points were used to isolate individual trees with a treetop detection algorithm and a tree boundary extraction algorithm. Lastly, tree species could be delineated in the hyperspectral images. Although this method was able to increase the accuracy of tree-level identification, it was computationally demanding and not scalable to larger spatial scales. Future work should focus on algorithm optimisation and computing hardware upgrade to improve data processing efficiency.
The high point density of airborne LiDAR coupled with the high spectral resolution of aerial imagery is an exciting approach for fine estimation. High-density 3D point clouds and high spatial and spectral resolutions together can ensure high estimation accuracy. More importantly, unlike satellite imagery, the quality of the fused data is not compromised by synoptic conditions, because the data are obtained via active sensors. However, expensive flight costs and time-consuming data processing limit its tractability for large-scale applications.

Theme 2: Multi-Temporal Input
Cross-sectional modelling of urban forest attributes and their ecosystem services might be biased by unobserved factors. A useful strategy to mitigate this bias is to replicate observations of the same units over time, thus time-invariant and location-specific unobserved factors can be controlled [72]. In particular, multi-temporal remotely sensed data carry great potential for capturing local details and have greater potential to improve estimation accuracy based on locally detailed and spatially explicit data input. For example, using multi-temporal image data can increase the success rate of mapping urban tree species composition as phenological information is retrieved [41,48]; this could be of vital importance for urban forests due to their compositional and configurational complexity and heterogeneity. Due to such potential, empirical studies have made use of short or long series multi-temporal remote sensing datasets to estimate changes in urban forest compositional and configurational attributes (e.g., dominant species, forest age, canopy density); hence, a reliable relationship between urban forests and the ecosystem services they provide is established.

Short Series Data Input
Some studies have utilised short series (usually only two) datasets to estimate changes in urban heat island effects and the urban forest structure during a given time span, especially before and after a certain event. Rogan et al. [73] integrated satellite and aerial images to examine the change in urban heat island effect induced by eradicating the invasive Asian Longhorned Beetle, which caused the elimination of 30,000 trees since 2008 in central urban areas in Massachusetts, USA. Landsat-5 (TM) thermal images acquired in the summertime in 2008 and 2010 were used to obtain land surface temperature. NAIP images also acquired in both years during leaf-on conditions were used to extract tree canopy cover. The authors found that the large amounts of tree loss and increased impervious surface between 2008 and 2010 were significant in determining the increased land surface temperature.
Similarly, Landsat-5 and -8 images from 2004 and 2014 were examined to detect changes in urban forest structure in Jinjiang city, southern China [74]. Combined together with ground observations and spatial statistical modelling, the mechanisms that influence the surface temperatures of urban forest canopies were successfully explored.
Qian et al. [75,76] utilised HSR and multispectral satellite images to explore the changes in urban greenspace configuration induced by the 2008 Olympic Games in Beijing, China. Taking advantage of both the HSR and multispectral features of SPOT and ALOS images taken in 2005 and 2008, fine-scale greenspace coverage and landscape indices were derived for spatial analysis. The authors found that holding the Olympic Games in Beijing not only incentivised the local government to increase greenspace coverage but also changed the spatial pattern of greenspaces across the city.

Long Series Data Input
The majority of recent studies uses multi-temporal data to study changes in urban forests and the induced socioeconomic and environmental effects based on long series data input (16 of 26). The ability of multi-temporal remote sensing to monitor long-term changes in urban vegetation cover, urban forest structure, and urban heat island is profound. Its potential also lies in monitoring changes at various scales, from a small urban park to a metropolitan city and even to an entire region or nation. Landsat imagery has been largely used by long-term studies owing to its long and continuous time-series. MODIS imagery is also popular in urban heat island studies because it has a superior temporal resolution than Landsat. The advantages of using multi-temporal satellite images for a long-term, multi-temporal study are apparent: low cost, instant access, and short revisit period.
Yang et al. [77] estimated the changes in urban greenspace coverage in 30 major Chinese cities at regional level between 1990 and 2010. Landsat-5 (TM) and Landsat-7 Enhanced Thematic Mapper Plus (ETM+) images acquired in 1990, 2000 and 2010 were used to classify land cover into five types with the SVM method. They found highly divergent trends of urban greenspace changes across different cities, which could be explained by varied developmental patterns and government policies.
Adam and Smith [78] explored the relationship between vegetation cover and urban heat island effect across the greater Sydney metropolitan region, in Australia. All available MODIS thermal band scenes (102 in total) with less than 5% cloud cover in the summertime between 2000 and 2011 were used as the land surface temperature index. The vegetation cover indices (e.g., foliage projective cover, fractional cover of photosynthetically active vegetation) were extracted based on Landsat imagery and ancillary data within the same time-series. All indices were resampled to 1 km 2 to match MODIS spatial resolution. Unlike most previous studies that employed single-phase remotely sensed data and a single mode of measuring vegetation cover, the multi-temporal satellite imagery employed in this study was capable of accounting for temporal variability and estimating the regional effects of varying density and structure of vegetation cover.
Bae and Ryu [79] examined the spatial and temporal variations of soil organic carbon stock in a constructed urban park in Seoul, Korea, and explored how soil organic carbon stock was shaped by land use and land cover changes. In their study, the NDVI values were derived from MODIS data. Landsat-7 and -8 time-series images acquired between 2003 and 2013 were collected to measure the temporal changes in NDVI values and land use pattern. Changes in the soil organic carbon concentration of topsoil were quantified based on land use types. The results indicated that land use history of the urban park played an important role in the temporal variations of the soil organic carbon concentrations.
Di Leo et al. [57] explored the role of urban green infrastructure in mitigating the urban heat island effect over 20 years in an urban area in sub-Saharan Africa. A total of 16 processed satellite images acquired from Landsat-5 (TM), Landsat-7 (ETM+) and Landsat-8 Operational Land Imager and Thermal Infrared Sensor (OLI/TIRS) between 1991 and 2013 were used to extract data on greenspace cover and land surface temperature. Data alignment was required for multi-temporal images from different Landsat sensors. On the one hand, the spatial resolutions of the thermal bands of the three Landsat sensors were different: 120 m for Landsat-5, 60 m for Landsat-7, and 100 m for Landsat-8. On the other hand, the thermal resolutions were also divergent: Landsat-5 and Landsat-7 had only one thermal band (Band 6), while Landsat-8 had two thermal bands (Band 10 and Band 11). This study demonstrates the advantages of multi-temporal satellite images in supporting spatiotemporal studies of urban forests and the urban heat island. However, the insufficient number of high-quality periodic satellite images for a 20-year analysis constitutes a critical limitation in this study.
Overall, for long-term high-frequency monitoring it is difficult to ensure that all images are cloud-free, despite the tools available on Earth Explorer and Google Earth Engine that allow researchers to select the most cloud-free scenes from the library. Sensor anomalies, as found with Landsat (https://www.usgs.gov/land-resources/nli/landsat/landsat-known-issues), may also impact long-term data collection. Accordingly, multi-source data are sometimes required to compensate the need for high-quality images.

Theme 3: Multi-Scale Input
The relationship between urban forests and ecosystem services has been shown to vary across space at different scales [72,80]. Indeed, one of the key advantages of remote sensing is its scalability. Correspondingly, a more recent theme in remote sensing application in urban forests is the emergence of multi-scale studies, expanding from micro-scale (e.g., individual tree and park patch) to meso-scale (e.g., city) and macro-scale (e.g., regional analysis), although relatively limited attention has been given to this aspect (eight of 63 studies). Multi-scale studies are complex to carry out and require a thorough knowledge of different sensors and algorithms for fusing multiple images.

Local-and City-Scales
Duncan et al. [72] assessed how urban vegetation configuration (grass, shrub and tree) influences urban temperatures in Perth, Australia. The authors used multi-temporal, very fine spatial (20-cm) and vertical (1-mm) resolution aerial imagery (provided by Urban MonitorTM from the Commonwealth Scientific and Industrial Research Organisation), as well as MODIS land surface temperature data (from the Terra satellite and Aqua satellite) for the 2003 to 2016 time period. In the study, local detailed and multiple time-period inputs were incorporated into urban climate modelling at local-and city-scales.
The benefit of urban forests in mitigating runoff in highly urbanised central Beijing (the capital city of China) was measured in another study using IKONOS images with four spectral bands captured in July 2012 [81]. In this study, spatial data pertaining to land cover types (including the tree canopy of an urban forest) were extracted and used as input of a hydrologic model to simulate various runoff scenarios from small sites to the entire study area.
More generally, a spatially explicit multi-scale indicator was developed to integrate social data (residents' perception of multiple ecosystem services provided by urban forests) with remote sensing data (from high-resolution satellite images and LiDAR) to understand how the provision of ecosystem services is distributed in a city and predict how people perceive their provision across the entire urban landscape [5]. In particular, this modelling process and the indicator proposed can be easily scaled up to large regional scale and beyond.

City-and Regional Scales
Due to the considerable spatial heterogeneity of urban forests across large areas, regional analysis of urban forests and associated ecosystem services using remote sensing techniques is still rarely conducted. Nevertheless, urban tree canopy cover in Los Angeles and Sacramento, California was mapped using QuickBird satellite and aerial imagery [82]. Combined with field survey data (e.g., age and species composition of urban trees), carbon storage and sequestration density were quantified across the study region.
At the national level, Landsat-5 (TM) and Landsat-7 (ETM+) SLC-on images with a processing level at Level 1T of 30 Chinese cities in 1990, 2000, and 2010 (obtained from the Earth Resources Observation and Science Center, USGS) were used to investigate changes in urban green spaces [77]. Combined with socioeconomic and natural data, the potential driving forces behind the identified changes were then explored.

Challenges and Future Directions
The volume of remote sensing images continues to grow at an enormous rate due to rapid advances in sensor technology and the launch of new satellite constellations. Consequently, an increasing amount of images from various sensors have become available, including multi-resolution and multi-temporal as well as multi-spectral band images. Despite this upsurge in trends, challenges have emerged in the field of urban forestry concerning the use of remote sensing techniques and the realisation of their full potential. First, although many remote sensing data and products are available (Table 3), there are technical issues with regard to data fusion and synergy (such as fusing optical, hyperspectral imagery, synthetic aperture radar remote sensing data that operate at disparate resolutions and use diverse acquisition approaches) as well as data matching to meet urban foresters' practical needs [35]. Second, a direct link is still lacking between ecological and geographical features, which can be extracted from remote sensing, and the adequate monitoring and modelling of ecosystem services provision, which usually requires diverse data inputs and sophisticated data processing [27]. Third, although several urban forest management software programmes (e.g., i-Tree developed by the USDA Forest Service) have standardised remote sensing data inputs, they are far from being developed as internationally transferable tools due to insufficient field calibration and validation [83]. And fourth, the techniques employed for urban tree/forest recognition, classification, segmentation and change detection from remote sensing data are still in their infancy, even though the key advantage of remote sensing technology lies in its capability of integrating data from multiple sources with multi-spatial/spectral/temporal resolutions that can provide enhanced data on urban forest resources and more robust and interpretable information for assessing ecosystem services.
In line with the aforementioned challenges, we present a set of future research directions on the application of remote sensing in an urban forest as follows: (1) There is a need for robust algorithms to fully automate the co-registration of data captured by multiple sensors as well as an optimal combined strategy of fusion algorithms to integrate multi-source remote sensing data at the pixel, feature and landscape levels.
(2) Another direction is to calibrate modelling techniques (such as processing accuracy and efficiency in dealing with spatial/temporal co-registration and high data dimensionality) to long-term time series so that past and present monitoring and analyses can evolve to innovative forecasting with regard to the dynamics of urban forests and their ecosystem services. Ultimately, temporal analysis would become a precursor to predicting the future and informing long-term urban forest strategy based on trends derived from past observations [84]. (3) Owing to the complexity of natural and anthropogenic factors in urban areas, urban forest studies relying on remote sensing data at a larger scale are still at the experimental stage [27,50].
Multi-scale analyses would help to access spatial information over a wide range of scales from local to regional and even global. Yet, scale variations should be carefully considered when using remote sensing applications to inform land-use policies and the related estimation of ecosystem services provision.

Concluding Remarks
In tandem with the ever-growing availability of high-resolution remotely sensed imagery from diverse airborne and spaceborne platforms, research continues to broaden the application and utilisation of remote sensing techniques to support and advise urban forest studies, such as the accurate modelling and quantification of urban ecosystem services [27,35]. To effectively and adequately realise the potential of remote sensing, a strong connection and collaboration between urban foresters and remote sensing experts is highly advised. This conjoint synergy would allow to recognise and address the aforementioned challenges as well as ensure the optimisation of urban forests and their ecosystem services for long-term human wellbeing.