Comparative performance of four machine learning models for land cover classification in a low-cost UAV ultra-high-resolution RGB-only orthomosaic

The necessity for accurate and affordable geospatial information continues to grow in modern times. For this purpose, we employed a low-cost unmanned aerial vehicle (UAV) equipped with a 1-inch Hasselblad CMOS sensor. The UAV captured ultra-high-resolution (UHR) images having a ground sampling distance of 3.7 cm. The overlapping raw images were processed with the structure for motion technique to produce an orthomosaic. We investigated the comparative performance of four machine learning (ML) models on this UHR RGB-only orthomosaic for multi-class land cover classification. The significance of this research lies in working without NIR bands as well as utilizing minimal training data. We evaluated support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), and maximum likelihood classification (MLC) to map seven land cover classes. Confusion matrices were constructed to thoroughly test the strengths and weaknesses of each ML model. SVM outperformed KNN, MLC, and RF with a high overall accuracy of 90% and a kappa coefficient value of 0.88. The non-parametric nature of SVM and its ability to handle smaller training samples were considered major contributing factors to its success. Additionally, the introduction of a specific class ‘asphalt with dirt’ facilitated the identification of road spectral variances. This effectively mapped road heterogeneity and contributed to the overall enhancement of results. In this research, we benefitted from the synthesis approach of working with multiple disciplines such as photogrammetric engineering, remote sensing, computer vision, and machine learning. This approach can propel the development of economical geospatial solutions.