Remote Sensing-Based Mineral Prospectivity Mapping Using Machine Learning

The prediction of mineral resources is a crucial aspect of mineral exploration,which is necessary to meet the growing demand for mineral resources in industrial development countries.Mineral prospectivity mapping(MPM)is a multi-variable decision-making tool that is used to identify and prioritize high-potential zones for mineral exploration in untapped regions.MPM relies on establishing a function that integrates geological features(input variables)with the presence of a targeted mineral(output variables).This integration is achieved by analyzing the spatial relationships between input features and known mineral occurrences using numerical algorithms.It is therefore critical to choose an appropriate algorithm that can accurately learn the complex relationships between variables in order to obtain a reliable prediction.The most important step in mineral prospectivity modeling is the selection of evidential features that represent the spatial manifestation of ore-controlling factors.The satellite multispectral remote sensors represent a cost-effective source of data for fast geological investigation.Although remote sensing data have the capability of detecting and delineating geological and structural features that aid in identifying new areas of mineralization,they have not been fully explored as means of identifying mineral exploration targets using the latest developments in machine learning predictive modeling.In this regard,this thesis systematically investigates the capabilities of three multispectral sensors data,namely Landsat-8,Sentinel-2,and ASTER,as well as data synergy,to map mineral prospectivity based on various machine learning and deep learning models.This thesis proposes a framework for exploiting remote sensing data as an alternative technique for MPM practice in case of data availability issues,while it provides a meaningful reference for detecting the likelihood of gold deposits in an orogenic mineralization environment.The main points of this thesis are as follows:(1)The current study describes the advantages of using remote sensing and available primary geological data to produce various thematic layers of ore controlling factors of west Hamissan,NE Sudan.These thematic layers are combined in a way that forms four distinct datasets of predictor variables,namely Landsat-8,Sentinel-2,ASTER,and data-integration datasets.Each dataset contains several layers,which are produced based on the spectral capability of the sensor to meet one of the exploration criteria of the orogenic gold deposits.(2)The spatial analysis methods,including Euclidean distance,and kernel density,were used to process the geological data and generate geological-based predictor maps.In the meantime,remote sensing-based predictor maps were produced based on different remote sensing enhancement techniques,including band ratio(BR),principal component analysis(PCA),and minimum noise fraction(MNF).These enhancement methods were employed to present the hydrothermal alteration zones related to gold deposits in the study area.By stacking these predictor layers including the geological-based maps and the remote sensing maps,12-,13-,and 24-layers form Landsat-8,Sentinel-2,and ASTER datasets,respectively,while the data-integration dataset consists of 41 layers.(3)These four datasets were used to train a set of machine learning algorithms,including random forest(RF),support vector machine(SVM),artificial neural network(ANN),and convolutional neural network(CNN).The three machine learning models and one deep learning model are employed to conduct data-driven gold mineral prospectivity modeling of the study area,based on(i)different models’ configuration of various sets of parameters,(ii)training each algorithm using each dataset separately,(iii)evaluating the results through accuracy assessment of the model classification and prediction performances.(4)The modeling results revealed that the ASTER dataset outperformed Sentinel-2 and Landsat-8 datasets.Whereas ANN and CNN achieved classification accuracies of 80%,meanwhile CNN yielded the best predictive performance based on(0.946)area under the ROC curve(AUC).By using the data-integration concept,the prediction accuracy of ANN and CNN models decreased about 2% and their classification accuracy increased by about 7% compared with the ASTER dataset.RF model trained by data-integration dataset achieved predictive performance(AUC:0.937)better than the other three models.Based on the fusion of very high gold zones of four models derived from the data-integration dataset,6 favorite tracts were suggested as worthy for further future geological exploration.