Change Detection in Complex Vegetation Communities

This project develops and demonstrates methods to introduce UAV sourced remote sensing products for monitoring and reporting change in complex vegetation communities. Multicamera rigs collecting single narrow spectral bands reconstructed the most complete tree canopy models while providing imagery for recording vegetation type.

Near Infrared and red edge imagery reconstruct tree and vegetation features while red band images reconstruct ground and near ground features. Imagery captured at 87% forward: 80% side lap and 80m above ground level generated 7-8cm resolution products that were able to detect mortality in branches of single shrubs while reconstructing tree canopies >35m above ground level. Geolocated field photos and transect intercept records are suitable for training and validation but location inaccuracy prevents use of quadrat based assessment. Dispersed ground reference marks (<150m) recorded at survey grade are required for monitoring at sub-decimetre resolutions (8cm).

The first complete and direct community boundary mapping using dense point clouds improved accuracy by 2-3fold with 90% of field ecological boundary locations within measurement error. Baseline dense point vegetation reconstruction varied by approximately 1% but by combining image sets collected at different times of day 3D vegetation reconstruction differences were <4% between seasons and it was possible to extract reproducible high resolution (0.16m GSD) canopy height models. Shadow and variable illumination affect reconstruction and classification of very high resolution imagery.

This project merged three normalised vegetation indices to reduce the impact of shadows on classification accuracy. Processed imagery and digital surface models were used for an ensemble of machine learning and statistical classification algorithms. Training data was constructed from air photo interpretation of high resolution RGB products supported by geolocated field photos. Industry standard (ArcGIS) and open source (R-GDAL) software workflow tools were explored and classification workflows validated using Neural Networks, Support Vector Machines, Decision Tree and Random Forest approaches.

A fully operational ArcGIS Toolbox using Machine Learning and Maximum Likelihood classifiers is delivered. The workflows report both accuracy and full confusion matrices for transparent comparison of methods. Cohen Kappa >0.9 were achieved for complex vegetation targets.

This report demonstrates the capability to detect between year changes in vegetation cover and tree canopy growth. Multiple epoch remote sensed imagery is essential to monitor complex communities. Bare ground, litter, senescent vegetation and water are simply classified and detection of Eucalyptus establishment is demonstrated.