Open Cut » Environment
The purpose of project work was to investigate colour and infrared aerial imaging for monitoring the revegetation of coal mining sites. Project duration (2006-2010) coincided with a period when wet emulsion aerial survey was replaced with direct digital capture of imagery and the automated extraction of terrain/surface models.
Aerial surveys of Saraji, Norwich Park and Gregory Coal Mines were conducted by AAM in 2006, 2007, 2009 and 2010. These were followed with detailed field observations by the ecological survey team (consultants Bevan Emmerton and James Elsol). Field data were used to determine the spectral signatures of ground cover types and for testing the accuracy of map classifications. Remotely sensed data were pre-processed, by AAM and Bluesky International (UK), to generate the spatially registered othophotos and terrain models necessary for image analysis work (also used to guide software development) at Buckinghamshire New University.
Techniques investigated at Saraji, Norwich Park and Gregory Coal Mines included: (1) conventional image classifiers; (2) advanced machine learning techniques; (3) spectral band transformations (vegetation indices, data reduction with principal and canonical components); (4) supplementing spectral data with expressions for vegetation height (derived from terrain and „canopy‟ surface models); and (5) the use of recently developed object-segmentation algorithms („watershed‟ techniques for spatial clustering of spectrally similar pixels).
Of all techniques trialed Maximum Likelihood (ML) classification in combination with both object-segmentation and vegetation height, consistently produced land-cover classifications that were highly correlated with ground-truth test data. In digital tests, a classification accuracy of 92% obtained for a 70Km2 site at Gregory Mine indicates that this technique has operational potential for monitoring broad classes of vegetation structure and condition over extensive areas.
Project results, therefore, suggest that the ML and certain machine learning techniques are sufficiently sensitive to monitor against key rehabilitation completion criteria, such as for the cover and slope requirements recommended by Grigg and co-workers (ACARP Project C8038; 2001) for pasture in the Bowen Basin.
More detailed ecological classification may be achieved with ground observation appropriate for extracting the digital signatures of desired cover classes. However, it is probable that such classifications may benefit from other technologies (such as LiDAR) to determine how spectral, pattern and structural characteristics of vegetation associations relate to remotely sensed signals.
Of other techniques, productive areas for further enquiry included: (1) that some machine learning approaches should be explored using training data more precisely aligned to „finer‟ cover classes; and (2) a combination of Canonical Correspondence Analysis (CCA) transformation and a fractal expression for the first Principle Component demonstrated sensitivity to dry and dormant vegetation.
Directions suggested for future research and operational work include ground, survey, technical and algorithmic methods for more detailed ecological classifications, improved techniques for recording vertical/canopy structure (including LiDAR). Operational measures for ensuring the reliability and repeatability of classifications include: avoidance of shadow and off-nadir perspective effects; continued development of workflow automation and distributed reporting mechanisms; and quality control throughout all stages of survey, orthomosaic and terrain/surface model production.