Multi-element quantitative extraction of mining area of laterite nickel mine combined deep learning with object-oriented method

Abstract

Remote sensing technology has great advantages in timely and rapid monitoring of open-pit mining conditions. Based on Worldview-2 multi-spectral satellite remote sensing images, we took a typical lateritic nickel deposit in Indonesia as an example, and 9 types elements which are mining area, dump, collection sump, tailings pond, buildings, roads, smelter, vegetation and bare soil in mining active areas were extracted. Firstly, a deep learning method based on TensorFlow framework was used to extract the main roads and mining areas from the pre-processed images to obtain vector data. Secondly, according to the vector data, our study area can be divided into two areas, center and outskirts, by FNEA coarse segmentation, and the local variance change rates of the two areas are calculated, so as to select appropriate segmentation scales for each factor type and establish a bottom-up multi-scale segmentation hierarchy. Thirdly, the spectral difference index (SDI) and PCA-based GLCM texture features were proposed to expand the feature base. The FSO algorithm and SEaTH algorithm were combined to select the optimal features and separation thresholds. At last, the multi-element extraction of laterite nickel ore area was completed hierarchically. The overall accuracy reached 90.12%. Our results indicated that the proposed method takes into account the spatial differences of various elements, ensuring the accuracy of segmentation and element extraction. Furthermore, method of selecting scales and thresholds avoids multiple experiments and reduces the time and labor cost of trial and error, which ensures objectivity and improves the selection efficiency. In addition, the PCA-based texture features can shorten the feature calculation time from 6.25 min to 14 s, reducing the operation time of the algorithm and greatly saving the operation time while ensuring the correlation effectively.