Exploring satellite image analysis methods for characterizing Canterbury shelterbelts, and the application to carbon modelling

Abstract

Shelterbelts are a prominent part of Canterbury’s agricultural landscape. Despite this, shelterbelts are not particularly well characterized in a spatially explicit manner. This study aimed to develop relatively automated methods for delineating and characterizing shelterbelts using high spatial resolution satellite images, and to then apply these techniques for the purpose of modelling shelterbelt carbon quantities. First, per-pixel and object-oriented image classification methods for delineating shelterbelts from QuickBird images (0.6 m) were compared. Object-oriented classification with Feature Analyst, a feature extraction software, was the most successful in delineating shelterbelts from the images, with an overall classification accuracy of 92 %. A statistical modellng method (Random Forests) was then investigated to determine its utility in differentiating shelterbelt tree species using spectral information. Shelterbelt data collected from three study areas (16 km² each) within the Hurunui District of Canterbury were used. Shelterbelts could be reliably differentiated into broad species groups to an accuracy of more than 90 %. However, differentiating individual coniferous shelterbelt species proved to be more challenging, with a model accuracy of only about 60 %. Results suggested that blue and red bands are important differentiators of broad species groups, whereas green and near infrared bands are important differentiators of individual coniferous species. Finally, shelterbelt carbon was modelled as an example application of shelterbelt delineation and species differentiation methods. Shelterbelt carbon was estimated using predetermined allometric equations which utilize field collected measurements. Field-based estimates were then used to model tree biomass using remotely-sensed variables as predictor variables. Regression analysis determined that shelterbelt species, the median value of the red band, shelterbelt width and tree density are all significant predictors of above ground biomass. Analysis results suggested that 2.6 % of land used for agriculture in the Canterbury Plains are comprised of shelterbelts. This study confirmed that P.radiata and C.macrocarpa are the major shelterbelt species in Canterbury, comprising 95 % and 3 % of shelterbelts, respectively. This study also suggests that shelterbelts represent a significant carbon reservoir on the Canterbury Plains, sequestering an average of 381 tonnes per hectare of shelterbelt. Currently, only major shelterbelts are accounted for by LUCAS (Land Use and Carbon Analysis System) in the Land Use, Land Use Change, and Forestry (LULUCF) Sector. This study has estimated that shelterbelts contribute at least 9.7 t/ha of carbon to the low producing grassland carbon pool, which is currently estimated at 29 t/ha. This study has demonstrated that shelterbelts can be successfully delineated and characterized using satellite images. With improvements, the methods presented in this study could be used in the future to semi-automatically delineate and characterize shelterbelts across large agricultural areas of New Zealand, such as the Canterbury Plains. The methods presented here are valuable tools for natural resource management. Potential applications include modelling the effectiveness of shelterbelts as wildlife corridors, assessing the change in shelterbelt landcover resulting from agricultural intensification, and quantifying the shelterbelt carbon pool across a given landscape.