Lithology classification in semi-arid area combining multi-source remote sensing images using support vector machine optimized by improved particle swarm algorithm

Abstract

The development of multi-source remote sensing technologies is helpful for geologists to obtain more comprehensive and complete lithological maps. In recent years, establishing automatic classification models based on Machine Learning (ML) algorithms has become an important approach to identify various lithologies supported by remote sensing data. Aiming at the specific geological and geographical conditions in a semi-arid area, Duolun County, Inner Mongolia Autonomous Region, China, this paper integrated GaoFen-2 (GF-2), Sentinel-2A, Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and GaoFen-3 (GF-3) remote sensing data, and used Support Vector Machine (SVM) classifier on the basis of Particle Swarm Optimization (PSO) to carry out the lithology classification. Firstly, on the basis of removing the interference of vegetation information from the acquired remote sensing data, a 63-dimensional candidate feature sequence was constructed by extracting spectral, backscattering, polarization and texture features. Secondly, an improved PSO algorithm with the Inertia Factor changing with the S-curve Decreasing function (SDIF-PSO) was proposed, and on this basis, a feature selection and lithology classification algorithm using SVM classifier based on two-layer SDIF-PSO was designed. Finally, the iterative optimization process of multiple optimization algorithms for SVM model parameters and the lithology classification accuracy before and after feature selection were compared. The experimental results showed that the proposed SDIF-PSO algorithm had the best optimization capability, with the highest cross-validation accuracy of 90.90%, which was improved by 3.85% than that of Grid-Search Optimization (GSO) algorithm, and 0.15% than that of the improved PSO algorithm with the Inertia Factor changing with the Linear Decreasing function (LDIF-PSO) and the improved PSO algorithm with the Inertia Factor Decreasing with the Concave function (CDIF-PSO). The dimension of the best feature combination was reduced to 35 through feature selection, and the convergence cross-validation accuracy reaches 92.14%, which was improved by 1.24% than that of all 63-dimensional candidate features in the same optimization process using SDIF-PSO algorithm.

Keywords

• Multi-source remote sensing
• Particle swarm optimization
• Feature selection
• Semi-arid area
• Lithology classification