Experiments Synergizing Machine Learning Approaches with Geospatial Big Data for Improved Urban Information Retrieval

Abstract

Geospatial big data produced by airborne or ground-based hyperspectral sensors have high spectral resolution and therefore account for the heterogeneity of urban areas. Machine learning approaches like support vector machine (SVM) successfully handle noise and variability in such datasets. SVM performs intermediate of spectral angle mapper (SAM) and object-based image analysis (OBIA) in retrieving buildings and natural features from the 2 m airborne hyperspectral data over Reno. However, such data acquisition is not economically viable for large scale detailed urban information extraction. Sensor optics design also forbids the exchanging of descriptive spectral and spatial content. Super-resolution (SR) reconstruction offers a solution. Using iterative back projection (IBP) and bicubic interpolation over the Washington DC scene shows that IBP creates a higher spatial resolution image at the same scaling ratio without retaining spectral characteristics. Algorithms sparse regression and natural prior (SRP), and anchored neighborhood regression (ANR) ensure input spectral information preservation while reconstructing the scene for Ahmedabad by learning the relationship between the spectral response and the feature patches. Visual examination and quality metric evaluation show that SRP outperforms ANR. SAM, SVM, and OBIA classify the best and worst super-resolved products to prepare urban material and land cover maps. A comparative assessment evinces that radial kernel-based SVM classifies the SRP generated output most favorably. These real data experiments thus establish SR as a tool for synthesizing spatially and spectrally rich datasets and highlight the flexibility of machine learning for geospatial big data tasks.