Quantification of Carbon Sequestration in Urban Forests

by   Levente Klein, et al.

Vegetation, trees in particular, sequester carbon by absorbing carbon dioxide from the atmosphere, however, the lack of efficient quantification methods of carbon stored in trees renders it difficult to track the process. Here we present an approach to estimate the carbon storage in trees based on fusing multispectral aerial imagery and LiDAR data to identify tree coverage, geometric shape, and tree species, which are crucial attributes in carbon storage quantification. We demonstrate that tree species information and their three-dimensional geometric shapes can be estimated from remote imagery in order to calculate the tree's biomass. Specifically, for Manhattan, New York City, we estimate a total of 52,000 tons of carbon sequestered in trees.



There are no comments yet.


page 2

page 3

page 4


Urban Tree Species Classification Using Aerial Imagery

Urban trees help regulate temperature, reduce energy consumption, improv...

Identification of Tree Species in Japanese Forests based on Aerial Photography and Deep Learning

Natural forests are complex ecosystems whose tree species distribution a...

From Google Maps to a Fine-Grained Catalog of Street trees

Up-to-date catalogs of the urban tree population are important for munic...

Species tree estimation using ASTRAL: how many genes are enough?

Species tree reconstruction from genomic data is increasingly performed ...

A Flexible Pipeline for the Optimization of CSG Trees

CSG trees are an intuitive, yet powerful technique for the representatio...

Ranked Schröder Trees

In biology, a phylogenetic tree is a tool to represent the evolutionary ...

Three-dimensional Segmentation of Trees Through a Flexible Multi-Class Graph Cut Algorithm (MCGC)

Developing a robust algorithm for automatic individual tree crown (ITC) ...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Environmental reports (IPCC, ) underline the pressing need for the elimination of Green House Gases (GHG) from the atmosphere in order to bring the carbon dioxide level to the pre-industrial norm. Carbon removing techniques span from scrubbing emission sources, manufacturing carbon trapping materials, and sequestering carbon in trees or soil. One popular idea proposed recently is the afforestation of 900 billion hectares of land (Bastin et al., 2019), which has the potential to offset more than 200 megatons of carbon from the atmosphere. In addition, in the emerging carbon trading market, companies may purchase forested land to offset their GHG emission and reduce their carbon footprints. There is a need for tools and platforms that can quantify and track GHG emission and total carbon offset. Such tools may need to estimate the total carbon stored in trees or in soil, multiple times per year, to ensure a fair and transparent carbon trading market.

Currently, carbon sequestration is estimated by a plethora of proprietary tools and models, making it hard to compare side by side carbon sequestration models. In general, carbon estimates rely on generic models where the shape, density, and species distribution of trees are surveyed from small sample plots and then extrapolated to large geographies and environments (Sileshi, 2014). The maximum amount of carbon that can be captured by a tree is mainly limited by the geometric sizes of trees which are fundamentally bounded by basic physics principles, like optimal water transport from roots to leaves, in addition to local soil and climate that drives growth rates (Koch et al., 2004). Therefore, knowledge of tree coverage, their geometric sizes and species characteristics are crucial in making good carbon storage estimates, but it is challenging since this information is not readily available for most of the locations on the planet (Chave et al., 2014).

In this work, we propose to use remote sensing data to determine tree coverage, estimate tree’s geometric shapes and species, and ultimately quantify the carbon sequestration in those trees at scale. Specifically, we train machine learning models to analyze aerial imagery to determine tree coverage, to classify their species, and to determine the local allometric relation for each tree species using only LiDAR data from a sample region to calibrate locally the biomass estimation model. Then the model is extended to nearby regions where LiDAR data may not be available but aerial imagery exists. As an illustration, we generate a high-resolution map of carbon sequestration in New York City urban forest.

Figure 1: Data processing workflow to quantify carbon sequestration in trees based on aerial images, LiDAR data, and tree species data.

2 Related work

Tree allometric estimation.

The allometric scaling relates tree height to tree crown diameter and tree dimensions to tree biomass. Such correlations are commonly done in forestry research (Chave et al., 2005). Estimates of the tree size are important to infer the total biomass and subsequently the carbon stored in trees. It is recognized that tree sizes are limited by water transportation, gravity, and environmental conditions (Koch et al., 2004). There is a quest to establish a generic scaling law correlating tree height with tree lateral dimensions (Klein et al., 2019) to better estimate the above ground biomass (Chave et al., 2014). Since the scaling relationship strongly depend on tree species and tree age in addition to the locations where trees are growing (urban vs rural areas), carbon estimate models need local calibration based on tree species and geographical conditions.

Tree species classification.

Quantifying the carbon sequestered in trees is limited by the exact knowledge of tree’s geometrical dimensions and their species. Current advances in image processing enable plot-level tree classification and estimates of tree sizes (Guan et al., 2015). Here we propose to generalize the image-based estimation method, to identify individual trees and to classify their species (Zhou and Klein, 2020) for urban forests using only aerial imagery and labeled tree species data.

Tree’s total biomass.

Estimates of carbon captured by trees are based on knowing the tree’s volume and the tree’s wood density that is strongly tree species dependent. In order to determine the volume, the tree crown diameter and tree height need to be determined in addition to tree species. The above ground biomass (AGB) can be estimated using a simple approximation (Chave et al., 2005):


where is tree height, denotes tree canopy diameter, specify the tree dry mass density, and is a form factor that takes into account tree shapes. The shape factor can vary from 0.01 to 1 depending on the tree trunk shape (Chave et al., 2014). The below ground biomass (BGB) is assumed as (Cairns et al., 1997) and the total biomass of a tree is specified by the sum of AGB and BGB.

3 Method

Our pipeline for quantification of carbon sequestration employ aerial imagery and LiDAR data. Figure 1

illustrates the data processing steps and the employed machine learning models to determine tree coverage, tree geometric sizes and species, and in the end, quantify the carbon sequestration. The main components are: (1) a Random Forest Classifier to identify tree-covered areas combined with image segmentation to delineate individual trees, (2) Deep Learning models to classify tree species, and (3) a carbon calculator tool to estimate the total biomass and carbon sequestered in trees.

3.1 Data

2D Imagery.

The National Agriculture Imagery Program (NAIP) is acquired every other year at a spatial resolution of 0.6 m. Multispectral bands of red, green, blue and near-infrared are simultaneously collected during full leaves season. NAIP is consistently collected for the last two decades, making the data source ideal to track tree coverage, tree growths, and to detect changes in land coverage.

Figure 2: Multispectral NAIP imagery (a) and the corresponding segmentation for individual tree crowns (b). The allometric relation between tree crown diameter and tree height for Pin oak tree is illustrated (c) and the estimates of its total biomass (d).


LiDAR 3D point clouds are used to extract tree heights and to calibrate the allometric relations for each tree species. Compared to aerial imagery, LiDAR data is much more expensive to collect and often unavailable. Therefore, it is critical to estimate tree height from 2D imagery to be able to scale to large geographies where only aerial imagery is readily available.

Tree Species.

Labeled tree species data is available from many municipalities (NYC-Street-Tree-Map, 2015) as part of their effort to quantify urban forest benefits (OpenTrees.org, ). Typically, data collection is crowd-sourced and specific tree attributes are captured like tree species, tree location, and diameter at breath height.

3.2 Tree detection and segmentation

Spectral information from remote sensing images has been used to delineate trees from other land cover classes. Using the red and near-infrared spectral information,we compute the Normalized Difference Vegetation Index (NDVI) (Pettorelli, 2013) for NAIP imagery. The NDVI is commonly used to separate vegetation from other classes like roads, buildings, bare land, and/or water. Within the vegetation class, further classification is achieved by training a Random Forest (RF) model to distinguish trees from grasses or bushes after incorporating additional information like image texture information. The texture is quantified within a sliding window of 5

5 pixels where the maximum, minimum, mean, average, and standard deviation are calculated for each or combination of spectral bands. The spectral and texture information is used by the RF classifier to identify areas that may be covered by trees.

Once the tree mask is generated, segmentation algorithms like “Watershed” (Wang et al., 2004) are used to cluster pixels that share common spectral and textural characteristics. The clustered pixels are converted to polygons to identify the tree crown boundaries and determine the canopy diameter. Tree crown diameters are then used to correlate against tree heights to establish the allometric equation for each tree species.

3.3 Tree height estimation

The height of the tree can be estimated from the crown diameters with a species-specific allometric equation. Here the allometric equation is modeled with a linear model111In some cases a more complex relationship may exist between tree crown diameter and tree height (Chave et al., 2005), and the linear model may over or under represent biomass estimates. mapping the crown size extracted from NAIP imagery to the tree height ground truth extracted from LiDAR data. The model can then be applied to areas where no LiDAR data is available.

3.4 Tree species classification

Since tree species information is crucial in estimating carbon storage, we train a convolutional neural network (CNN) to classify tree species from NAIP imagery. In our approach, the NAIP images are diced into 32

32 tiles and all four bands of the NAIP imagery are used compared to standard models where only three bands, RGB, are commonly used. The network is a modified ResNet34 (He et al., 2016) which allows four-channel images as input instead of three channels. The training data are prepared by cropping the NAIP data around each location of the labeled tree species data. Then the trained model is run across target areas to generate a tree species map.

3.5 Carbon sequestration

Carbon stored in trees is considered to be around 50% of the total biomass of a tree (Thomas and Martin, 2012). The AGB can be calculated for each tree based on crown diameter and the corresponding height estimation. Based on Section 2, the carbon sequestered in trees can be estimated as .

4 Experiments and Results

Figure 3: Tree species classification map for four tree species in New York City. The corresponding labels are referenced in Table 1.

We apply our pipeline to the New York City (NYC) area to demonstrate the process of carbon sequestration quantification. LiDAR data was acquired in 2017 over Staten Island borough, NYC (NYC-LiDAR, ), and tree species data was collected across all five boroughs in NYC in 2015 along the public roads only (NYC-Street-Tree-Map, 2015). Trees that were not near roads are not included in the survey as well as trees in parks or private properties. A total of 56 NAIP tiles (50 km by 50 km in size, 50 GB data in total) are processed. A sample view of the RGB composite is shown in Figure 2a.

For tree delineation, an RF classifier using two classes, tree and anything else, is used to delineate trees from any other land cover classes. Training data are manually labeled and are used to train the RF model that generates a tree coverage map. Once the tree-covered area is separated from other land cover classes, individual trees are delineated using the watershed segmentation method as shown in Figure 2b. The tree crown diameter is then calculated based on the assumption that the crown is circular in shape and diameter is proportional to the square root of the polygon area.

Tree type Label # points (kg/m)
London planetree 0 55,903 560
Honeylocust 1 43,974 755
Callery pear 2 42,384 690
Pin oak 3 30,575 705
Total 172,876
Table 1: Dataset of tree species and tree dry mass density.

The same NAIP tiles are used for tree species classification. The dominating four species in NYC are sampled to generate the training data, as listed in Table 1. The model achieves a classification accuracy of 80% on the test split. The model is then applied across all the NAIP data after splitting into 3232 tiles except the ones with a mean NDVI value lower than 0.0 (typically non-vegetation areas like buildings, roads, waterbody, etc. (Zhou et al., 2020)). A tree species classification map is shown in Figure 3. The tree species of each tree crown polygon from tree delineation is then determined by majority voting.

The next step is to estimate tree height from crown diameters. We first process the LiDAR point cloud to generate a canopy height model which is a height-from-ground map. For each crown polygon, multiple points were queried against the LiDAR height map and the mean value was calculated as tree height ground truth. Linear regression between the tree crown diameter and tree height is calculated for each tree species for a training data set from the LiDAR covered area. The linear regression curves (shown in Figure 

2c) are used to infer the tree height from the crown delineating polygons.

The above ground biomass is calculated for each tree based on crown diameter and tree height taking into account the tree-specific density as listed in Table 1 (World-Agroforestry, ). The form factor for each tree is , same for all tree species, and the AGB is estimated based on Equation 1. The total biomass is dependent on tree sizes, as shown in Figure 2d for Pin oak.

Figure 4: Carbon stored in individual trees mapped on the Upper West Side of NYC.

A map of carbon sequestered in trees is shown in Figure 4 for the Upper West Side of the New York City area with the lower-left corner of the image showing Central Park. The total carbon stored in the urban forest in Manhattan borough, NYC is estimated to be 52,000 tons, obtained by summing up the individual tree carbon storage. In comparison, the total carbon stored in NYC urban forest is estimated to be 1.2 million tons by Nowak et al. (Nowak et al., 2018), and the carbon stored in Manhattan trees is 43,500 tons based on the ratio of the number of trees in Manhattan to the total in NYC, which is consistent with our estimation.

5 Conclusions

Quantifying tree’s carbon sequestration can enable a better carbon trading marketplace where the information can be shared and trusted. Here we demonstrated such an approach built on public data sets where aerial imagery and tree label data can be used to build high-resolution carbon sequestration maps. The method allows to map carbon sequestered at individual tree level and spatially aggregate the carbon to city level.

Broader Impact

Carbon trading markets and GHG offset require transparent and verifiable methods to quantify the total carbon sequestration. The “bottom-up” approach introduced in this work, can estimate the total carbon sequestered in trees and spatially map the carbon stored in urban and rural forests and track year-to-year changes in carbon sequestration.


  • J.F. Bastin, Y. Finegold, C. Garcia, D. Mollicone, M. Rezende, D. Routh, C.M. Zohner, and T.W. C. et al (2019) The global tree restoration potential. Science 365 (6648), pp. 76–79. Cited by: §1.
  • M. A. Cairns, S. Brown, E. H. Helmer, and G. A. Baumgardner (1997) Root biomass allocation in the world’s upland forests. Oecologia 111 (1), pp. 1–11. Cited by: §2.
  • J. Chave, C. Andalo, S. Brown, and et al. (2005) Tree allometry and improved estimation of carbon stocks and balance in tropical forests. Oecologia 145 (1), pp. 87–99. Cited by: §2, §2, footnote 1.
  • J. Chave, M. Réjou‐Mécha, A.. Búrquez, E. Chidumayo, M.S. Colgan, W.B. Delitti, and et al. (2014) Improved allometric models to estimate the aboveground biomass of tropical trees. Global change biology 20 (10), pp. 3177–3190. Cited by: §1, §2, §2.
  • H. Guan, Y. Yu, Z. Ji, J. Li, and Q. Zhang (2015) Deep learning-based tree classification using mobile lidar data. Remote Sensing Letters 6 (11), pp. 864–873. Cited by: §2.
  • K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In

    Proceedings of the IEEE conference on computer vision and pattern recognition

    pp. 770–778. Cited by: §3.4.
  • [7] IPCC The intergovernmental panel on climate change. External Links: Link Cited by: §1.
  • L. J. Klein, C. M. Albrecht, W. Zhou, C. Siebenschuh, S. Pankanti, H. F. Hamann, and S. Lu (2019) N-dimensional geospatial data and analytics for critical infrastructure risk assessment. In IEEE International Conference on Big Data (Big Data), pp. 5637–5643. Cited by: §2.
  • G. W. Koch, S. C. Sillett, G. M. Jennings, and S. D. Davis (2004) The limits to tree height. Nature 428 (6985), pp. 851–854. Cited by: §1, §2.
  • D.J. Nowak, A. R. Bodine, and et al. (2018) The urban forest of new york city. Technical report Technical Report 117, Vol. 82, Newtown Square, PA: US Department of Agriculture, Forest Service, Northern Research Station. Cited by: §4.
  • [11] NYC-LiDAR Topobathymetric lidar data (2017). External Links: Link Cited by: §4.
  • NYC-Street-Tree-Map (2015) New york city street tree map. Technical report New York City, NY. External Links: Link Cited by: §3.1, §4.
  • [13] OpenTrees.org OpenTrees. External Links: Link Cited by: §3.1.
  • N. Pettorelli (2013) The normalized difference vegetation index. Oxford University Press. Cited by: §3.2.
  • G.W. Sileshi (2014) A critical review of forest biomass estimation models, common mistakes and corrective measures. Forest Ecology and Management 329 (), pp. 237–254. Cited by: §1.
  • S.C. Thomas and A.R. Martin (2012) Carbon content of tree tissues: a synthesis. Forests 3 (2), pp. 332–352. Cited by: §3.5.
  • L. Wang, P. Gong, and G. S. Biging (2004) Individual tree-crown delineation and treetop detection in high-spatial-resolution aerial imagery. Photogrammetric Engineering & Remote Sensin 70 (3), pp. 351–357. Cited by: §3.2.
  • [18] World-Agroforestry ICRAF database. External Links: Link Cited by: §4.
  • W. Zhou, L. J. Klein, and S. Lu (2020) Pairs autogeo: an automated machine learning framework for massive geospatial data. In IEEE International Conference on Big Data (Big Data), pp. . Cited by: §4.
  • W. Zhou and L. J. Klein (2020) Monitoring the impact of wildfires on tree species with deep learning. arXiv:2011.02514, pp. . Cited by: §2.