Geostatistical simulation for filling misclassified pixels in hyperspectral images

• Published in: AIP conference proceedings Vol. 2209; no. 1
• Main Authors: Varga, Sebastian, Benndorf, Jörg
• Format: Journal Article Conference Proceeding
• Language: English
• Published: Melville American Institute of Physics 25.03.2020
• Subjects: Algorithms; Computer simulation; Face recognition; Geological mapping; Geostatistics; Hyperspectral imaging; Optical measuring instruments; Outcrops; Pixels; Research projects; Underground mines; Underground mining
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0000282

Mine of the future or mining 4.0 is the subject of numerous current research projects. There are many research questions to answer. One important aspect is to investigate the possibility of an automated geological detection and face mapping, especially in underground mining. Optical sensors are used for automatic detection tasks in industrial sorting. Especially hyperspectral sensors are widely used to identify materials and are therefore able to identify the ore. This can be shown in various studies on the mapping of outcrops in open pits. To use this technique in an underground mine requires the development of a method that can cope with the challenges of this complex environment including inconsistent light conditions, high variability of the mineral distribution and associated reflection issues. Due to these environmental conditions, poor data quality must be expected at least in some image sections. As a result, many pixels cannot be classified correctly. To solve this problem, this contribution proposes in a first step to identify these not uniquely classified pixels. The image is classified using fuzzy C-means and the poorly classified pixels are identified. In a second step, an attempt is made to determine a possible class affiliation for these identified pixels by analyzing neighborhood conditions using geostatistics. In particular, possible realizations for these poorly classified pixels are simulated using the Single Normal Equation (SNESIM) algorithm. The combination of the classification action results with the geostatistical simulation using Bayes statistics leads to the final results. This paper presents an investigative study of this proposed approach and discusses results.