Rapid assessment of mine rehabilitation areas with airborne LiDAR and deep learning: bauxite strip mining in Queensland, Australia

Abstract

Rehabilitation of disturbed sites is a complex process requiring constant monitoring and assessment, with current methods being somewhat obsolete, manual, and rather fraught with risk in terms of beneficial outcomes of the evaluation. This study challenges the status quo of manual assessments and build a new methodology to recognise a clear need for a quasi-automated and reliable remote sensing method to achieve rapid mine rehabilitation monitoring. The aim of this study was to innovate and develop an ensemble analytic methodology for the use of airborne LiDAR, in conjunction with deep learning, to rapidly evaluate the rehabilitation status of post-mined areas. Airborne LiDAR datasets, obtained via an Optech ORION m/c 200 sensor, were collected over bauxite mining operations on the western coast of Cape York Peninsula, Queensland, Australia. These datasets were processed into a 3 D point cloud using the Esri ArcGIS Pro platform. An ensemble analytical approach was developed, leading with a Convolutional Neural Network (CNN) algorithm, which was trained to recognise, then classify images of the reference site (natural ecosystems) and active rehabilitation areas. The result of this classification was then leveraged, training in turn a support vector machine (SVM) algorithm which assessed the entire LiDAR collection, resulting in a predictive surface of rehabilitation status. This surface, as the final analytical product, was assessed for validity using ground-truth plot data and an ordinary least squares regression model, as well as comparison to subject matter expert evaluation through a Delphi expert panel assessment. The results indicate that the predictive surface, as the final analytical product, had statistically significant correlation coefficient with multiple attributes (or predictor) that correlated with the rehabilitation status interpretation. These predictor variables were primarily structural in nature and included the canopy height, sub-canopy height, key species as a proportion of canopy, herbaceous ground cover percentage, litter ground cover percentage, and distance from haul road (p < 0.05). A Delphi expert collaboration comprised of ten environmental science subject matter experts and 600 observations, indicated a high precision (r = 0.89 at p < 0.05) of the prediction surface when compared with a subject expert panels assessment, as well as a high Total Deviation Index (TDI) of 16% (p < 0.05). This indicated positive agreement between the panel’s assessment and the predictive surface. This study suggests that airborne LiDAR and deep learning ensemble analytics can be used as a rapid means to estimate the rehabilitation status of disturbed sites.