Abstract. A back-propagation artificial neural network (ANN) model is pro- posed to discriminate zones of high mineral potential in the Rodalquilar gold.
. . , 2003, . 24, . 5, 1151–1156
Artificial neural networks as a tool for mineral potential mapping with GIS J. P. RIGOL-SANCHEZ Department of Geology, University of Jaen, 23071, Jaen, Spain
M. CHICA-OLMO and F. ABARCA-HERNANDEZ RSGIS Lab, Department of Geodynamics, University of Granada/IACT, Avda. Fuentenueva s/n, 18071, Granada, Spain (Received 9 July 2001; in final form 14 August 2002) Abstract. A back-propagation artificial neural network (ANN) model is proposed to discriminate zones of high mineral potential in the Rodalquilar gold field, south-east Spain, using remote sensing and mineral exploration data stored in a GIS database. A neural network model with three hidden units was selected by means of the k-fold cross-validation method. The trained network estimated a gold potential map efficiently, indicating that both previously known and unknown potentially mineralized areas can be detected. These initial results suggest that ANN can be an effective tool for mineral exploration spatial data modelling.
1. Introduction Mineral exploration is a complex task which often requires the use of remote sensing, geochemical, geophysical and geological data to achieve satisfactory results. This demands a multithematic approach, and has made geographical information systems (GIS) a popular tool among exploration geologists because they allow for more effective integration and analysis of large numbers of spatial data with different attributes and formats (Burrough and McDonnell 1998). Data integration methods in GIS may be divided into two main groups based on knowledge-driven and datadriven models. Knowledge-driven models, such as weighted index overlay, fuzzy logic and multi-criteria evaluation, estimate the parameters of a function for combining datasets on the basis of expert opinion; while data-driven techniques such as regression, weights of evidence modelling, Dempster-Shafer belief functions, certainty factors, Bayesian statistics and artificial neural networks (ANNs), estimate model parameters from measured data (Bonham-Carter 1994). Zhou and Civco (1996) identified several problems affecting knowledge-driven data integration models, such as the difficulties inherent in handling spatial data that present inaccuracy and factor interdependency, assigning scores and selecting the aggregation function. Data-driven procedures also involve assumptions that are difficult to satisfy when dealing with geological variables such as linear relationships, variable independence and normal distributions; ANNs, however, are an exception and seem to offer advantages over other methods because they make no assumptions about the data and allow for International Journal of Remote Sensing ISSN 0143-1161 print/ISSN 1366-5901 online © 2003 Taylor & Francis Ltd http://www.tandf.co.uk/journals DOI: 10.1080/0143116021000031791
1152
J. P. Rigol-Sanchez et al.
nonlinear and interactive effects (Bishop 1995). For instance, Singer and Kouda (1996) have shown the potential of ANNs in the search for mineral deposits using X-ray mineralogical data from 69 drill holes. Although ANNs are used frequently as remote sensing data classifiers, their potential as spatial modelling tools is still underexplored. One reason for this might be that ANNs are often regarded as ‘black-boxes’ along with the fact that ANN modelling involves an appreciably more elaborate training process (i.e. parameter adjustment) than other methods. In the present study, a back-propagation neural network model was developed using a comprehensive exploration GIS database for mineral potential mapping in the Rodalquilar gold mining district, Spain (figure 1(a)). A gold potential map was generated with the model. 2. Study area and datasets The test site in Almerı´a, south-east Spain, covering an area of 150 km2, corresponds to the Rodalquilar mining district (figure 1(a)), where epithermal quartz-alunite gold deposits have been exploited in recent decades. They are associated with felsic to intermediate tertiary volcanic rocks showing fracturation and pervasive hydrothermal alteration (Rytuba et al. 1990). On the basis of the deposit model for the district outlined by Rytuba et al. (1990), the following datasets were selected for the analyses: (a) Landsat 5 Thematic Mapper (TM) image acquired in July 1991; (b) geochemical survey (59 elements, 372 locations); (c) gravity and magnetic survey (330 ground stations); (d) fracturation map; and (e) gold occurrence database containing 49 locations. Datasets were gathered in the context of the DARSTIMEX European Project (Chica-Olmo et al. 2002), where a discussion about the simulation of different mineral exploration scenarios using weighted index overlay, is presented. TM data processing was focused on identifying hydrothermally altered areas. TM band ratios have been shown to be very useful in delineating rock alteration in the area (Crosta and Moore 1989). TM 5/7 and 3/1 band ratios were selected as input variables to the model because of their ability to discriminate ore-related hydroxyl and iron-oxide alteration respectively. Geochemical data were processed by performing principal component analysis (PCA) on 46 selected mineralization-related elements. PC1 and PC3, indicating lithology (Si-Al versus Ca) and Au-Ag-As-S association respectively, were chosen for further modelling. Continuous surfaces were created by kriging (Journel and Huijbregts 1978) from geochemical PCs to minimize estimation error. Gravity and magnetic residual values were also interpolated by kriging to generate residual anomaly maps indicating the presence of potential ore-related buried anomalous bodies. A distance-to-nearest-fracture map was derived from the fracturation map by using GIS analysis functions. All seven input maps were stored as raster layers with 30 m spatial resolution in the exploration GIS (figures 1(b–h)). Preliminary statistical analyses indicated that input variables showed non-normal distributions and cross-correlations (Spearman R) between 0.10 and 0.55. 3. Methodology 3.1. T raining set construction A feed-forward network was used for this study (see Bishop 1995 for further details). Training patterns were created by recording the values shown by input GIS variables at each of the 46 locations of the gold occurrence database. The training dataset was completed adding 59 locations scattered over the district selected by means of stratified random sampling and recording input variable values as above.
Remote Sensing L etters
1153
Figure 1. (a) The study area, and input layers: (b) distance to nearest fracture; (c) TM 5/7 band ratio; (d) TM 3/1 band ratio; (e) geochemistry PC1; (f ) geochemistry PC3; (g) gravimetric anomaly; (h) magnetic anomaly. Lighter tones indicate higher values.
The reason for this was twofold: to include both positive (presence) and negative (absence) patterns, and to reflect, to some extent, the positive skew probability distribution of known occurrences in the district. It was therefore assumed that the prior probability of the 59 points containing a deposit was very low. Thus, gold occurrences were coded as 1 and the remaining locations as 0. These values constituted the desired output. Input variables were linearly rescaled to the range [0, 1]. As discussed below, ANNs can be developed even when the pattern set is small.
1154
J. P. Rigol-Sanchez et al.
3.2. Network implementation The network implementation required three main steps: model and architecture selection; training; and performance assessment. First, the pattern set was partitioned randomly into a training set (86%) and a test set (14%) for independent testing (figure 2). Since the numbers of input and output layer units were set to seven and one respectively, architecture selection consisted of choosing the number of hidden layer units. This step was performed using k-fold cross-validation on the training set, which is a method that is especially well suited to small pattern set modelling (see Bishop 1995). The training set was randomly divided into k=10 subsets of equal size. Different networks with increasing number of hidden units were then trained k=10 times, each time leaving out one of the subsets from training, but using only the omitted (accuracy assessment or, in ANN parlance, validation) subset to assess their generalization performance (figure 2). In all experiments training was stopped when no further decrease in error on the accuracy assessment subset (i.e. generalization) occurred over 200 cycles. A variant of the back-propagation algorithm with momentum and a flat spot elimination term was used for training. The activation function was taken to be the logistic function and the learning rate was set to 0.4. The training process was run six times using different random initial weights. The model having the smallest average accuracy assessment error over 10 subsets (60 runs) was selected. Independent generalization performance assessment was done by presenting the network with the previously withheld test set with no learning (i.e. training) taking place and computing its average error. This ensured that the model will not only fit the training data, but also previously unseen data, which is essential in predictive mapping. Finally, the
Figure 2. Schematic view of network model implementation procedure.
Remote Sensing L etters
1155
selected model was trained on the complete training set using the test set to stop training. The number of training cycles and the previously-derived average test error were also used to guide the process. A mineral potential surface (map) was generated by presenting the trained network with an input pattern set from the entire district. During this phase (i.e. estimation) the output unit generated a value within the range 0–1, which can be interpreted as the favourability (probability) to the presence of gold occurrence at a given location. 4. Results, discussion and conclusions Cross-validation indicated that the simplest network architecture yielding the smallest generalization error was one with three hidden layer units (figure 3). While adding more hidden units decreased the training error, it made a negligible difference to the accuracy assessment error. It also helps to prevent overfitting, an undesirable situation where a model that is too complex fits both the data and the noise in the pattern set. The average accuracy assessment rms error was 0.298 and the average rms error for the independent test set was 0.364 (R2=0.50). The testing accuracy obtained was 80% using a cutoff value of 0.5. In the final training step, the rms error for the training set was 0.415 and its R2 value was 0.52, while the average rms error for the accuracy assessment set (in this case the test set) was 0.340 and its R2 value was 0.45. Training and accuracy assessment set accuracies were 86.7% and 73.3% respectively. These results suggest that the model could be used to estimate the presence/absence occurrence patterns bearing in mind the complexity of the phenomenon being modelled. The gold potential map for the Rodalquilar district generated by the network model in the estimation phase is shown in figure 4. Training and test set patterns are also shown. There is a similarity between the potential map and input layer PC3 (figure 1(f )), suggesting that variable PC3 has a notable influence on favourability. Main high potential areas are located around Cerro del Cinto (A), where extremely high-grade ore bodies have been exploited, Los Tollos (B), Las Nin˜as (C), and north of B (D). Most known occurrences are adequately detected by the model, although some discrepancies can be noticed. The model failed to detect the test set occurrence labelled 1 and the training set occurrences west of location 5. On the other hand, previously unknown areas such as zone D and the two zones on the western edge are clearly delineated by the model, suggesting additional potential targets for further exploration. Initial results of this study demonstrate that ANNs can be used effectively as
Figure 3. Structure of the network model used for gold potential mapping.
1156
Remote Sensing L etters
Figure 4. Gold favourability map estimated by means of the 7-3-1 model. Training set and test set locations are shown as black and white markers respectively (dots=0, crosses=1).
tools for mineral potential mapping using remotely sensed and field exploration data. Research is currently in progress to evaluate a range of data integration procedures for mineral potential mapping. References B, C. M., 1995, Neural networks for pattern recognition. 1st edition (Oxford: Clarendon). B-C, G. F., 1994, Geographic Information Systems for geoscientists: modelling with GIS. 1st edition (Ottawa: Pergamon). B, P. A., and MD, R. A., 1998, Principles of Geographical Information Systems. (Oxford: Oxford University Press). C-O, M., A, F. and R, J. P., 2002, Development of a decision support system based on remote sensing and GIS techniques for gold-rich area identification in SE Spain. International Journal of Remote Sensing, 23, 4801–4814. C, A. P., and M, J. M., 1989, Geological mapping using Landsat Thematic Mapper imagery in Almeria Province, south-east Spain. International Journal of Remote Sensing, 10, 505–514. J, A. G., and H, C. J., 1978. Mining Geostatistics. (New York: Academic Press). R, J. J., A J., A., C, C. G., M, E. H., P, M. H., S, J. G., K, W. C., and A, A., 1990, Mineralized and unmineralized calderas in Spain, Part II: Evolution of the Rodalquilar caldera complex and associated gold-alunite deposits. Mineralium Deposita, 25 (Suppl ), 29–35. S, D. A., and K, R., 1996, Application of a feedforward neural network in the search for Kuroko deposits in the Hokuroku district, Japan. Mathematical Geology, 28, 1017–1023. Z, J., and C, D. L., 1996, Using genetic learning neural networks for spatial decision making in GIS. Photogrammetric Engineering and Remote Sensing, 62, 1287–1295.
