Towards lithology mapping in semi-arid areas using time-series Landsat-8 data

Abstract

Previous remote sensing studies have used optical images to map lithology, commonly based on a single-date product to extract spectral diagnostic features or other geology-related features to aid interpretation. Yet, the performance of time-series or multi-date optical images in lithology mapping has rarely been discussed. In this study, we employed time-series (TS) Landsat-8 products to map the lithology in a semi-arid area of Xinjiang, China. First, we extracted TS reflectance from Landsat-8 multispectral data, and then extracted TS surface moisture, greenness, and brightness via tasseled cap transformation (TCT), a method for transforming the spectral bands of optical images into components (brightness, greenness, and moisture) that can be physically interpreted. Moreover, the land surface temperature (LST) was directly extracted from Landsat-8 Collection 2 products. In total, five surface parameters (TS reflectance, TS moisture, TS greenness, TS brightness, and TS LST), were collected. Then, the rank-based non-parametric Kruskal Wallis rank-sum statistical test was applied, which demonstrated a significant difference between the TS surface parameters among different rock units. Furthermore, different combinations of the TS surface parameters were stacked separately and served as different input features for the random forest classifier. Finally, the performance of TS surface parameters in mapping lithology was evaluated. The classification results showed that (1) TS reflectance outperformed single-date reflectance for mapping rock units; (2) the combination of TS brightness, greenness, and moisture gave a comparable classification result, relative to TS reflectance; (3) TS LST performed best when only one single surface parameter was used to map lithology, and (4) the combined use of LST and all three TCT components achieved the highest accuracy (85.26%) with a kappa coefficient of 0.77. Although surface parameters derived from rock units changed inconspicuously or irregularly over time, these variations captured by TS Landsat-8 data enabled the improvement of the classification accuracy. Overall, our study illustrates the great potential and benefits of using TS Lansat-8 data to map rock units in arid domains.