Spatio-Temporal analysis with machine learning for sustainable management of abandoned quarries

Abstract

The abandoned quarries demand not only appropriate rehabilitation but also continuous monitoring. If these sites are left unmanaged, they can create significant environmental and ecological impact due to changes in land-use and land-cover. Monitoring of abandoned quarries are often overlooked due to challenges such as accessibility, safety and costs involved. Hence, there hardly exists a systematic monitoring approach or appropriate guidelines for managing quarry sites upon termination of the extraction activities. Thereby the presence of hazardous environments will be unavoidable with a substantial resistance on the quarry industry for not so sustainable closure procedures and following up actions. To overcome these challenges, sufficient monitoring of quarry sites and surroundings to enforce appropriate rehabilitation strategies with post monitoring that has minimal on-site involvements would be essential. For such purposes, analysing remotely sensed data would be applicable, based on the quality of data collection, processing and sufficient ground truthing. Accordingly, this study aims to develop an automated classification approach for mapping the land cover in the regions of abandoned quarries. It employs a comprehensive methodology that includes data preparation, feature extraction and selection, hyperparameter optimization, and identification of the algorithm that exhibits better accuracy. The efficacy of machine learning models - decision tree (DT), random forest (RF), and support vector machine (SVM) - were critical to analyse Landsat 8 and Sentinel 2 satellite images at selected sites in Anuradhapura, Sri Lanka. The outcome reveals that the SVM model produced the highest accuracy of 91.30% with a kappa index of 0.898. This superior performance of Sentinel 2 images could be attributed to their higher spatial resolution compared to Landsat 8 and SVM’s efficient handling of high-dimensional data. Furthermore, SVM’s robustness against overfitting using regularization, and its flexibility in dealing with complex separations through kernel functions would have facilitated the computations in addition to textural features and spectral indices incorporated to augment the model training procedure. It was also evident that augmenting the number of features can help alleviate the misclassifications that occur when exclusive use of spectral data. Utilizing the developed machine learning algorithms, a temporal analysis was performed on land cover from 2018 to 2022 to obtain a comprehensive overview on the changes. This analysis underscores the potential for monitoring land cover changes in abandoned quarries for effective management and rehabilitation strategies with minimal human intervention.