Geocoding of trees from street addresses and street-level images

https://doi.org/10.1016/j.isprsjprs.2020.02.001Get rights and content

Abstract

We introduce an approach for updating older tree inventories with geographic coordinates using street-level panorama images and a global optimization framework for tree instance matching. Geolocations of trees in inventories until the early 2000s where recorded using street addresses whereas newer inventories use GPS. Our method retrofits older inventories with geographic coordinates to allow connecting them with newer inventories to facilitate long-term studies on tree mortality etc. What makes this problem challenging is the different number of trees per street address, the heterogeneous appearance of different tree instances in the images, ambiguous tree positions if viewed from multiple images and occlusions. To solve this assignment problem, we (i) detect trees in Google street-view panoramas using deep learning, (ii) combine multi-view detections per tree into a single representation, (iii) and match detected trees with given trees per street address with a global optimization approach. Experiments for >50000 trees in 5 cities in California, USA, show that we are able to assign geographic coordinates to 38% of the street trees, which is a good starting point for long-term studies on the ecosystem services value of street trees at large scale.

Introduction

An urban forest inventory represents a “snapshot in time” of trees in a given geographic area (Roman et al., 2013). Urban tree inventories are gaining in popularity and increasingly entering the digital age, with digital devices replacing index cards, and data posted on websites facilitating public access and management (Twardus and Nowak, 2018). Of particular interest to municipalities are street tree inventories since street trees lie within their management scope and are a publicly visible component of the urban forest. Street tree inventories may be used for directing management priorities and assessing pest and pathogen risk (Bond, 2013), surveying species diversity and size distribution (McPherson and Kotow, 2013), and estimating ecosystem services (McPherson et al., 2005, McPherson et al., 2016). When updated over time, inventories can become key components of long-term monitoring data, which are essential for understanding changes in urban forests (Roman et al., 2013). Repeated inventories have been used by researchers to explore trends in street tree populations such as changes in species composition (Dawson and Khawaja, 1985), tree health impacts of storms (Hallet et al., 2018), estimates of demographic rates (Roman et al., 2013, Roman et al., 2014, Widney et al., 2016), as well as determinants of mortality or growth (Koeser et al., 2013, van Doorn and McPherson, 2018). Analyses of repeated tree measurements may help predicting population stability and identify trends. Baseline estimates of demographic rates such as mortality or replacement rates provide a basis for comparison to future measurements. Having information on how the urban forest is changing allows managers to identify the more vulnerable segments of the tree population and to focus management efforts. For example, areas of higher mortality can imply a need for maintenance, increased replacement planting, or a change in the species palette. Spatially-explicit tree data are important for several reasons (Alonzo et al., 2016): ecosystem functions may vary throughout a city due to legacy effects (Roman et al., 2017), species may have different threats from pests, diseases, and fire (Santamour, 1990, Laćan and McBride, 2008), and distribution of ecosystem services may be contributing to environmental injustices (Landry and Chakraborty, 2009). Geographic coordinates are key pieces of information for long-term monitoring studies because they allow for the same tree to be tracked and measured through time with some degree of certainty (van Doorn et al., in press).

For modern inventory collection, tress are typically mapped with GPS devices and thus come with a fairly accurate geographic coordinate, making repeated measurements much easier. However, before the prevalence and ease of GPS, trees in inventories were typically only referenced by street address. Manually linking such tree inventories to GPS-measured inventories is infeasible at a large scale. As a result, a lot of valuable inventory information remains unexplored because there is no automated way, yet, to link those two types of inventories.

Our main motivation is to provide a tool for building long-term tree data sets that can be used to assess changes in the urban forest and corresponding changes in ecosystem services. Although street trees make up a small proportion of the urban forest, they provide considerable ecosystem services (McPherson et al., 2016), and disservices to the urban landscape (Escobedo et al., 2011, Pataki et al., 2011). Benefits include improvement in air quality, a reduction of the heat island effect, increased carbon capture and storage, rising property values, and an improvement in individual and community wellbeing (Nowak et al., 2002, McPherson et al., 2016). According to the most recent estimate (McPherson et al., 2016) there are 9.1 million trees lining the streets of California with an ecosystem services value of $1 billion per year or $111 per tree, i.e. $29 per inhabitant of California. However, little information is available on how street tree populations are changing in terms of demographic rates. While static data sets such as canopy cover and plot data can provide a snapshot in time, demographic details such as growth rates and mortality rates, as well as changes in tree health through time require assessment of the same trees more than once. Many municipalities do not have the funding to complete repeat inventories and/or may place little priority on matching individual trees to create a time series. Thus, making accessible the information in legacy inventories is of high interest to both researchers interested in quantifying tree demographic rates and ecosystem services and for urban forest practitioners interested in data-driven management (e.g., knowing where there are hotspots of tree mortality, or high growth).

The hope is that our approach on retrofitting existing street tree inventories with geographic coordinates will enable large-scale longitudinal studies, where data of the same population of trees can be analysed over decades. Emphasis of our approach is thus on assigning geographic coordinates to a high absolute number of trees across the state of California as opposed to geocoding each individual tree of the given data base correctly.

In order to geo-code hundreds of thousands of street trees across California with corresponding addresses, we analyse publicly available street-level panoramas across the whole state. For this purpose, we propose to start from a simplified, computationally more efficient version of (Wegner et al., 2016, Branson et al., 2018) for tree detection and geo-localization. We replace the complex and computationally costly conditional random field formulation with a weighted average that condenses multiple detections of the same tree from different views into a single one. Tree instances detected in images are matched to inventory entries that come with street addresses. Preparing and training the system basically consists of manually labeling trees in images (four days), training the tree detector (four hours), and restructuring all inventories into a homogeneous format (one day). Our system comes at virtually no cost if we put aside costs for running computers and downloading Google images. Given appropriate hardware and a fast internet connection for downloading images, the method scales to arbitrarily large data sets.

Section snippets

Related work

Municipial tree inventories are traditionally collected by urban forestry professionals (either municipal staff or contracted-out tree management company arborists) but more recently, some programs also have significant citizen science components (Cozad et al., 2005, Bloniarz and Ryan, 1996, Roman et al., 2017). Variables collected in tree inventories vary by program goal, but the common minimum dataset typically includes tree species, condition, and location (Östberg, 2013, van Doorn et al.,

Methods

In this section, we describe our method to assign geographic coordinates to street trees given street addresses and ground-level panorama images. Given a data base of trees with street addresses, our new system presented here does the following (see flowchart in Fig. 1):

  • It retrieves approximate geographic coordinates for each address,

  • downloads the available street-view panorama images for each address,

  • automatically detects all trees per image,

  • integrates individual tree detections per image into

Results and discussion

We ran experiments with tree inventories from five cities in the state of California, USA, that come with both accurate geocoordinates and street addresses, which allows us to train, validate, and test our approach. The total number of input trees with street addresses is 57938 and the number of street trees per city varied between 599 (Brentwood) and 34585 (Palo Alto). The total number of trees per municipality is reported in the top row of Table 1.

All street tree inventories were originally

Conclusion

We have presented a novel approach to assign geographic coordinates to street-trees given street addresses and street-view panorama images. Enhancing existing tree inventories with geographical locations allows to automatically match inventories from different dates and to track trees across longer time spans. Depending on the information that is being recorded in tree inventories (e.g., presence/absence, trunk diameter, vigor), repeated measures make it possible to calculate tree mortality

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This project was supported by funding from the Hasler Foundation, the US Department of Agriculture-Forest Service, and the Swiss National Science Foundation scientific exchange grant IZSEZ0_185641.

References (65)

  • L.A. Roman et al.

    Data quality in citizen science urban tree inventories

    Urban Forest. Urban Greening

    (2017)
  • P. Sidike et al.

    dPEN: deep Progressively Expanded Network for mapping of heterogeneous agricultural landscape using WorldView-3 imagery

    Remote Sens. Environ.

    (2019)
  • N. van Doorn et al.

    Demographic trends in Claremont California’s street tree population

    Urban Forest. Urban Greening

    (2018)
  • J. Zhang et al.

    A hybrid framework for single tree detection from airborne laser scanning data: A case study in temperate mature coniferous forests in Ontario, Canada

    ISPRS J. Photogramm. Remote Sens.

    (2014)
  • Aval, J., Fabre, S., Zenou, E., Sheeren, D., Fauvel, M., Briottet, X., 2019. object-based fusion for urban tree species...
  • D. Bloniarz et al.

    The use of volunteer initiatives in conducting urban forest resource inventories

    J. Arboric.

    (1996)
  • Bond, J., 2013. Best management practices - tree inventories. In: 2nd ed., International Society of Arboriculture, pp....
  • R. Cao et al.

    Integrating aerial and street view images for urban land use classification

    Remote Sens.

    (2018)
  • X. Chen et al.

    Multi-view 3d object detection network for autonomous driving

  • City of Ithaca, 2018. Tree Inventory/GIS. https://www.cityofithaca.org/253/Tree-Inventory-GIS. Online; accessed...
  • M. Cordts et al.

    The cityscapes dataset for semantic urban scene understanding

  • Cozad, S., McPherson, G., Harding, J.A., 2005. STRATUM Case Study Evaluation in Minneapolis, Minnesota. Center for...
  • C. Crown et al.

    Every tree counts: reflections on NYC’s third volunteer street tree inventory

    Arboricul. Urban Forest.

    (2018)
  • J. Dawson et al.

    Change in street-tree composition of two. Urbana, Illinois neighborhoods after fifty years: 1932–1982

    J. Arboric.

    (1985)
  • A. Geiger et al.

    Vision meets robotics: The kitti dataset

    Int. J. Robot. Res.

    (2013)
  • R. Hallet et al.

    Assessing the tree health impacts of salt water flooding in coastal cities: A case study in New York City

    Landscape Urban Plann.

    (2018)
  • Hartling, S., Sagan, V., Sidike, P., Maimaitijiang, M., Carron, J., 2019. Urban tree species classification using a...
  • H. Kaartinen et al.

    An international comparison of individual tree detection and extraction using airborne laser scanning

    Remote Sens.

    (2012)
  • V. Krylov et al.

    Automatic discovery and geotagging of objects from street view imagery

    Remote Sens.

    (2018)
  • V.A. Krylov et al.

    Object geolocation using mrf based multi-sensor fusion

  • D. Kwak et al.

    Detection of individual trees and estimation of tree height using lidar data

    J. Forest Res.

    (2017)
  • F. Lafarge et al.

    Creating large-scale city models from 3D-point clouds: a robust approach with hybrid representation

    Int. J. Comput. Vision

    (2012)
  • Cited by (0)

    View full text