CONVOLUTIONAL NEURAL NETWORK AND LANDSAT 8 IMAGERY TO DETECT GOSSANS IN THE CANADIAN ARCTIC

Supergene weathering and oxidation of massive sulphide deposits often lead to the formation of iron caps, known as gossans, between the surface and the water table. Depending on the relative mineral abundances, the surficial expression of gossans is characterized by red to yellow hues. They are of interest to geologists because they constitute vectors to economic mineral deposits. Also, reactive gossans in permafrost are natural analogs for acid drainage in mine waste deposits.

Some gossans have already been studied from a remote sensing perspective. Their identification is generally possible by the observation of absorption features in their reflectance spectra. As gossans are Fe-oxides/hydroxides rich, the main absorption peaks are caused by electronic processes (Fe2+ and Fe3+), and vibrational processes (Fe-OH bonds). Mapping Fe rich formation or gossans has been conducted for over 4 decades using multispectral imagery and techniques such as bands ratios, principal components analysis or support vector machine.

However, a test of these methods to detect gossans mapped in various locations across the Canadian Arctic yielded a very low success rate, possibly because the previous methods are not applicable to the variety of gossaneous compositions and cannot account for the presence of sparse vegetation. Hence, in this study, we propose a new method to identify gossans and overcome these issues based on a convolutional neural network (CNN) approach and Landsat 8 data. A CNN is trained for binary classification (presence or absence of gossan) of Landsat 8 imagery. A wide variety of hyperparameters was tested. The model was trained using the locations of 809 gossans previously identified and mapped in the Canadian Arctic.

Our best model obtained an accuracy of 77 % for the detection of gossans on the test dataset (≈ 140 occurrences). The spectra of the 809 gossans were extracted from the imagery and compared with those extracted from randomly placed points. The figure shows that gossans, in our region of study, are almost spectrally indistinguishable from random points proving how challenging the detection is. For this reason, achieving 77 % of accuracy is a success and the method deserves further work. The use of higher spatial and spectral resolution imagery could provide more relevant information that could be associated with gossans.