GSA Connects 2021 in Portland, Oregon

Paper No. 245-15
Presentation Time: 5:05 PM


MARMOLEJO, José1, MCMILLAN, Nancy1 and BUENEMANN, Michaela2, (1)Department of Geological Sciences, New Mexico State University, Las Cruces, NM 88003, (2)Department of Geography, New Mexico State University, Las Cruces, NM 88003

Given the extent of mining activity in New Mexico's recent history, it is of interest to both government and general public to remediate hazards associated with active and abandoned mines. Extraction practices can release toxic substances into the environment. Furthermore, acid mine drainage, which is caused by the oxidation of mining waste, is a significant detrimental effect caused by improper mining practices. These adverse effects, when improperly managed, can result in large-scale contamination of local ecosystem and public health hazards (Pohl, 2011; Dalton et al., 2005). Thus, it is essential to develop techniques to effectively identify the locations of abandoned mines. The research uses abandoned mines in the Tortugas Mountains (TM) east of Las Cruces New Mexico to develop a method for rapid identification of abandoned mines.

The current method, according to the New Mexico Mining and Minerals Division (NMMMD) is through in-situ field campaigns. We propose Machine Learning (ML) modeling, specifically Support Vector Machine classifiers (SVC) trained on high-resolution multispectral World-View 3 (WV3) satellite imagery to automate the detection process.

SVC modeling is based on 86 different iterations, where each iteration contains identical training and testing pixels associated with mining activity and randomized variations of non-mining pixels to reduce ML bias effects. The SVC uses a training dataset with corresponding labels, to produce training vectors that are embedded into higher dimensional space using a radial basis function (RBF) kernel. A linear separating hyperplane is then created to segment the data based on the radial influence of vectors which is then verified using testing datasets. To find optimal solutions for this study, we perform a cross validation (CV) experiment on an individual iteration using 169 different SVC models to test all allowable permutations of hyperparameters.

Using CV optimized solutions, the learned hyperparameters are applied to all other iterations to validate parameter effectiveness. We report a 90.1 percent classification accuracy for abandoned mines.

Future work will use this SVC trained model to fit Stevenson-Bennet, Memphis, Morgan Manganese, and Rincon Manganese WV3 satellite datasets east of TM to assess model effectiveness for mine detection.