A GEOBIA framework to estimate forest parameters from lidar transects, Quickbird imagery and machine learning: A case study in Quebec, Canada

https://doi.org/10.1016/j.jag.2011.05.010Get rights and content

Abstract

The GEOgraphic Object-Based Image Analysis (GEOBIA) paradigm continues to prove its efficacy in remote sensing image analysis by providing tools which emulate human perception and combine analyst's experience with meaningful image-objects. However, challenges remain in the evolution of this new paradigm as sophisticated methods attempt to deliver on the goal of automated geo-intelligence (i.e., geospatial content within context) from geospatial sources. In order to generate geo-intelligence from a forest scene, this article introduces a GEOBIA framework to estimate canopy height, above-ground biomass (AGB) and volume by combining lidar (light detection and ranging) transects, Quickbird imagery and machine learning algorithms. This framework is comprised three main components: (i) image-object extraction, (ii) lidar transect selection, and (iii) forest parameter generalization. The rational for integrating these methods is to provide a semi-automatic GEOBIA approach from which detailed forest information is obtained at the individual tree crown or small tree cluster level (i.e., mean object size of 0.04 ha); while also dramatically reducing airborne lidar data acquisition costs. Analysis is performed over a 16,330 ha forested study site in Quebec, Canada. Forest parameter estimation results derived from our GEOBIA framework demonstrate a strong relationship with those using the full lidar cover; where the highest estimates for canopy height (R = 0.85; RMSE = 3.37 m), AGB (R = 0.85; RMSE = 39.48 Mg/ha) and volume (R = 0.85; RMSE = 52.59 m3/ha) were achieved using a lidar transect sample representing only 7.6% of the total study area.

Highlights

► We develop a GEOBIA framework to generate geo-intelligence from a forest scene. ► The framework includes image-object extraction, lidar transect selection and forest parameter generalization. ► Canopy height, biomass and volume generated from sampled lidar transects in this framework are highly correlated with those using the full lidar cover.

Introduction

Remote sensing techniques allow for the collection of Earth surface information over a range of scales in a synoptic and timely manner (Wulder, 1998). Today, high spatial resolution (i.e., H-res pixels generally less than or equal to 5.0 m) remote sensing data are rapidly accessible from a variety of sources, such as satellite-based optical sensors and airborne lidar (light detection and ranging) systems. Over the last decade, the development of new image processing techniques increasingly referred to as GEOgraphic Object-Based Image Analysis (GEOBIA) (Hay and Castilla, 2008) have proven effective for analyzing high resolution data by incorporating analyst's experience, complimentary ancillary data, sophisticated geospatial analysis and methods that emulate the human perception of image-objects within a scene (i.e., based on size, shape, tone, color, texture, topology and context), rather than as isolated pixels of varying color (Hay and Castilla, 2008, Blaschke, 2010). However, the evolution of GEOBIA faces a growing challenge to develop semi/automated methods that bridge the gaps between straightforward segmentation – the extraction of image-objects – and the generation of geo-intelligence from geospatial sources. Here geo-intelligence refers to geospatial content within context (Hay and Blaschke, 2010).

As a dominant terrestrial sink for atmospheric CO2, forests play an important role in the dynamics of the carbon cycle (Eamus and Jarvis, 1989). Similarly, precise forest management requires an accurate estimation of carbon content with an emphasis on above-ground biomass (AGB). To assess the commercial value of forests, volume is widely used to measure wood quantity; and an important parameter used to calculate AGB and volume, is canopy height. However, monitoring large-area forest parameters such as canopy height, AGB and volume, requires considerations of both accuracy and budget. Previous studies have proven promising to apply optical imagery and GEOBIA to retrieve forest parameters, such as forest height (Wulder et al., 2007, Mora et al., 2010), AGB (Addink et al., 2007, Kajisa et al., 2009), and volume (Mäkelä and Pekkarinen, 2001, Pekkarinen, 2002). Although it is cost effective estimating these parameters using only optical imagery, model accuracies are lower than those using airborne lidar data. To meet these challenges, recent research describes the combination of small-area lidar transects and wider extent optical imagery to provide cost-effective solutions. This is achieved by generalizing lidar-measured vertical canopy information from transects to the entire study site covered by an optical image (Hudak et al., 2002, Wulder and Seemann, 2003, Hilker et al., 2008, Stojanova et al., 2010, Chen and Hay, 2011). Recent studies (Chen and Hay, 2011, Chen and Hay, in press) have noted, that the ability to accurately extract this information depends on (i) the type of forest characteristics assessed, (ii) the ability to define appropriate lidar transects and (iii) the type of modeling and generalization methods used to relate transect samples back to the full scene.

Based on this brief background, the primary objective of this study is to present a GEOBIA framework to generate new forest geo-intelligence by estimating canopy height, AGB and volume from Quickbird imagery and airborne lidar transects. This framework builds upon prior research by incorporating three main components: (i) image-object extraction, (ii) lidar transect selection, and (iii) forest parameter generalization. Chen and Hay, 2011, Chen and Hay, in press first describe the use of a lidar transect selection algorithm and a support vector regression (SVR) generalization technique applied to a small (2601 ha) homogenous forest site (with two major tree species) in British Columbia, Canada. In this study, we build on this early work by presenting a more complete GEOBIA framework composed of one additional machine learning algorithm, and examine its performance over a larger (16,330 ha) more complex mixed forest site (with six major tree species), located in Quebec, Canada.

Section snippets

Study area

Our 16,330 ha (14.2 km × 11.5 km) study site (48°30′N, 79°22′W) is located in the Training and Research Forest of Lake Duparquet (TRFLD), Quebec, Canada (Fig. 1), where it is characterized as a South-Eastern Boreal Forest composed of an abundance of mixed stands. The site is dominated by balsam fir (Abiesbalsamea L. [Mill.]), along with white spruce (Piceaglauca [Moench] Voss), black spruce (Piceamariana [Mill] B.S.P.), white birch (Betulapaprifera [Marsh.]), trembling aspen (Populustremuloides

Data analysis

An important component of this project involves using optical imagery to generate ‘pseudo-height’ classes, from which to guide our selection and ‘acquisition’ of airborne lidar transects. This is based on research which describes useful relationships between optical imagery and canopy height (Franklin and McDermid, 1993, Hyde et al., 2006, Donoghue and Watt, 2006, Mora et al., 2010). Once transects are defined, forest height and species information (from the optical and lidar data covered by

Image-objects

Fig. 3(a) represents a sample area in our study site, covered by deciduous trees, conifers, roads and forest gaps. Fig. 3(b) shows the corresponding area overlaid by image-object boundaries, derived from the segmentation procedure. Fig. 3(c) represents an object-based image, where the spectral values within each image-object are averaged. We note that most image-objects in this figure have jagged boundaries, which are distinctly different from the boundary delineation results from many other

Conclusions

In this study, we have generated geo-intelligence from a forest scene by reducing airborne lidar data acquisition costs, providing meaningful geospatial information related to the size, orientation and location to best acquire lidar transects, and have applied novel machine learning algorithms to model important forest parameters over a large area. A semi-automatic GEOBIA framework is presented to extract forest information (i.e., canopy height, AGB and volume) at the small crown/cluster level

Acknowledgments

This research has been funded by an Alberta Informatics Circle of Research Excellence (iCore) Ph.D. scholarship awarded to Gang Chen. Dr. Hay acknowledges support from a Natural Sciences and Engineering Research Council (NSERC) Discovery Grant and an AIF New Faculty Award. Dr. Benoît St-Onge acknowledges support from the BIOCAP Foundation of Canada. We also thank the anonymous reviewers for their valuable suggestions.

References (45)

  • M.A. Wulder et al.

    Integrating profiling LIDAR with Landsat data for regional boreal forest canopy attribute estimation and change characterization

    Remote Sensing of Environment

    (2007)
  • X. Xie et al.

    Spacebased estimation of moisture transport in marine atmosphere using support vector regression

    Remote Sensing of Environment

    (2008)
  • E.A. Addink et al.

    The importance of scale in object-based mapping of vegetation parameters with hyperspectral imagery

    Photogrammetric Engineering and Remote Sensing

    (2007)
  • G. Camps-Valls et al.

    Robust support vector regression for biophysical variable estimation from remotely sensed images

    IEEE Geoscience and Remote Sensing Letters

    (2006)
  • G. Castilla et al.

    Size-constrained region merging (SCRM): an automated delineation tool for assisted photointerpretation

    Photogrammetric Engineering and Remote Sensing

    (2008)
  • Chang, C.-C., Lin, C.-J., 2001. LIBSVM: A Library for Support Vector Machines. Software available at...
  • Chen, G., Hay, G.J., Castilla, G., St-Onge, B., Powers, R. A multiscale geographic object-based image analysis (GEOBIA)...
  • Chen, G., Hay, G.J. A support vector regression approach to estimate forest biophysical parameters at the object level...
  • N. Cristianini et al.

    An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods

    (2000)
  • D.N.M. Donoghue et al.

    Using LiDAR to compare forest height estimates from IKONOS and Landsat ETM+ data in Sitka spruce plantation forests

    International Journal of Remote Sensing

    (2006)
  • S.E. Franklin et al.

    Empirical relations between digital SPOT HRV and CASI spectral response and lodgepole pine (Pinuscontorta) forest stand parameters

    International Journal of Remote Sensing

    (1993)
  • Gunn, S.R., 1998. Support Vector Machines for Classification and Regression. Technical Report,...
  • Cited by (64)

    • Forest height estimation combining single-polarization tomographic and PolSAR data

      2023, International Journal of Applied Earth Observation and Geoinformation
    • Above-ground biomass estimation from LiDAR data using random forest algorithms

      2022, Journal of Computational Science
      Citation Excerpt :

      The usefulness of LiDAR data in estimating forest characteristics has been largely proven [54], insofar as biomass has been estimated with very good results in different regions and with different species [55–57]. Although multivariate regression is the most popular approach, the complex relationships between forest variables are not always well captured by the models [58]. In order to overcome this disadvantage, and because the relevant literature reports better accuracy than linear regression techniques for biomass estimation [59], the potential of the RF estimation technique was thus the one researched in this study.

    • ML-LUM: A system for land use mapping by machine learning algorithms

      2019, Journal of Computer Languages
      Citation Excerpt :

      The GEOBIA is effective in remote sensing image analysis. However, it still faces challenges in order to achieve the goal of geospatial source automatic geographic intelligence [6]. Random Trees are an ensemble of Decision Tree (DT) classifiers.

    • Integrating multi-sensor remote sensing and species distribution modeling to map the spread of emerging forest disease and tree mortality

      2019, Remote Sensing of Environment
      Citation Excerpt :

      This map was later used to extract a large number of samples of sudden oak death distribution for reliable SDM calibration and validation (Section 4.4). To generate the map, we applied a geographic object-based Image Analysis (GEOBIA) framework following Chen et al. (2012). Compared to the classic pixel-based approach, GEOBIA uses image-objects (i.e., pixel clusters) as the basic study units to reduce errors caused by spectral variation within each geographic object (e.g., individual trees containing sunlit and shaded crowns; Chen et al., 2015a).

    • Estimation of forest structural and compositional variables using ALS data and multi-seasonal satellite imagery

      2019, International Journal of Applied Earth Observation and Geoinformation
    View all citing articles on Scopus
    View full text