b'AEGC 2023Short abstractsPortal, which continues to improve the veracity of petroleumenhancement problems in the computer vision field, but to date system modelling in Australian onshore basins. have received little attention in geophysics. This research uses GANs to enhance low-resolution magnetics data. Thirty thousand In summarising avenues for further work, the Onshore Basinpairs of high- and low-resolution images are constructed from Inventories programme has provided scientific and strategicexisting magnetic datasets in Australia. The dataset is split into direction for pre-competitive data acquisition under the EFTFtraining and validation sets and the training tiles are input into a work programme. Here, we provide an overview of the currentGAN model for training. The GAN model attempts to predict the status of the Onshore Basin Inventories, with emphasis on itshigh-resolution images from the low-resolution input, effectively utility in shaping EFTF data acquisition and analysis, as well aslearning the geophysical and spatial characteristics of the data, new gap-filling data acquisition. and the transformation between resolutions.This new GAN-driven resolution enhancement model AI/ML to unlock potential of dormant data of rock chips. demonstrates the potential of this technique to be used on Roman Beloborodov, Arun Sagotra, Marina Pervukhina,low-resolution datasets, particularly in areas where investment Richard Kempton, Valeriya Shulakova, Claudio Delle Piane, Juergin high-resolution datasets is limited. GANs have wide-ranging Hauser, Michael Ben Clennell and Lena Hancock applications to minerals exploration, groundwater studies, and planetary research.CSIRODrill cuttings or rock chips are an essential and frequentlyAn example of country scale airborneunderutilised source of subsurface information, often representinggeophysicsAngola.the only physical data recovered from a well. Using spectral and optical image data this study seeks to lithologically classify theJohn Bell and Billy Steenkamprock chips, estimate the per lithology mineral composition andA regional airborne geophysical survey was carried out over thereby implicitly identify the boundaries between differentAngola as part of a wider programme of investment and lithological units. The data are acquired using the HyLogger -regulatory changes to the countrys mining industry. Survey rapid spectroscopic logging and imaging system that collectsspecifications were lines 1 km apart with control lines spaced at hyperspectral measurements in the visible and infrared range as10 km. Total survey size was 1.4 million line km and was carried well as high-resolution optical images. Joint analysis of opticalout between 2014 and 2018. The nominal survey clearance was images and hyperspectral measurement on rock chips has the100 m. All surveys were flown with fixed-wing aircraft equipped potential to improve classification and mineral compositionwith horizontal gradiometers and gamma -ray spectrometers.estimation as the two type of measurements are complementary. Optical images provide information on texture and grain sizeThe countrys regional geological setting is examined in the distribution, while spectral data are sensitive to the mineralcontext of the airborne geophysical results. Angola is largely composition. The optical images are analysed with computerunderlain by the Congo Craton. The shield rocks are mostly vision and deep learning approaches to detect, segment,obscured by younger cover rocks. The location and extent of the and classify individual rock chips using pre-trained neuralLufilian Arc is unclear although tentative correlations are made networks. Spectral data are analysed using machine learning forwith the Damara Belt in Namibia. The geophysical signatures unsupervised classification, relying on prior geological knowledgeof the potash deposits of Cabinda are reviewed as well as the about mineral compositions of different lithotypes. The resultsregional setting of the Lucapa kimberlite field.of these analyses are then combined in a probabilistic model to improve lithological classification accuracy. Identifying, detecting,Several countries in Africa have completed or are in the process and characterising the individual rock chips provides criticalof acquiring regional surveys (Togo, Nigeria, Burkina Faso, etc). information for interpreting the lithological class and estimatingThis highlights the recognition that investment in the geological the mineralogical composition. The proposed approach has broadendowment is critical to attract investment in the mining applications, including but not limited to quantifying mineralindustry.fractions of energy-critical minerals in rock chips, iron ore fraction in iron ore tailings and ditch cuttings lithology quantification. Designing seismic surveys for reduced environmental impact.MAG-GAN: A deep learning, image super resolutionAlyson Birce, Andrea Crook, Peter Vermeulen, Mostafa technique for aeromagnetic data. Naghizadeh, Stephanie Ross and Cameron CrookAnthony Benn OptiSeis Solutions Ltd.Macquarie University Exploration and production projects, whether for oil and gas, Magnetic data are one of the most common geophysicalmining, or clean tech applications such as carbon capture and techniques in mineral and resources exploration. Developedstorage, typically begin with the acquisition of seismic data. nations have repositories of high-resolution geophysicalGenerally, these surveys involve optimising the geometry datasets. However, these resources do not exist for manyfor subsurface imaging, surface constraints, and operational regions worldwide. Consequently, there is demand for newefficiencies. However, an equally important aspect is to optimise methodologies of image enhancement, to capitalise onthe geometries for reduced environmental impact. This can existing low-resolution datasets. be accomplished by utilising alternative geometries or even a combination of geometries within a single survey. This paper Generative Adversarial Networks (GANs) are a type of deepfocuses on how to minimise the impact of seismic surveys on the learning algorithm that have been used fruitfully on imageenvironment while reducing costs and maintaining or improving FEBRUARY 2023 PREVIEW 82'