Incorporating Spatially-Explicit Uncertainty Metrics in Image-Derived Classification of Impervious Surfaces

Suny College Of Environmental Science And Forestry, Syracuse NY

Investigators

Abstract

Human-constructed impervious surface area (ISA) is an important indicator of human alterations to natural environments. ISA includes sidewalks, parking lots, driveways, and rooftops. In this project, the current state-of-the-art methods will be improved with a novel image analysis approach for detecting impervious surface areas. Typical single-thread multispectral classification will be extended into a hierarchical multi-process context-specific classification. This context-specific multi-process approach uses a variety of targeted machine learning algorithms while segmenting the problem into potentially easier sub-problems. These algorithms are selected based on the complexity of each underlying task. Therefore, they: i) are not restricted to a single machine learning method (e.g. decision trees, neural networks), and ii) support combination of various on-demand inputs. In addition to high classification accuracy, this hierarchical methodology allows identification of problematic cases through association of the end product with spatially-explicit uncertainty metrics. This approach is not constrained to ISA detection, but could be extended to other multi-spectral classification problems (e.g. vegetation, soil). Human-constructed impervious surface area reduces or eliminates the capacity of the underlying soil to absorb water. It has a significant impact on the environment and human health, therefore playing an important role in land use decisions. For example, impervious surfaces dramatically increase peak discharges associated with storm and snowmelt events, increasing the likelihood of downstream flooding as storm waters exceed stream channel capacities. This impervious surface area detection model will provide accurate results associated with advanced uncertainty metrics. The novel machine learning approach will enhance image classification algorithms, thus advancing science and engineering. Furthermore, the unique monitoring results are targeted for scientists not necessarily familiar with image classification and its limitations (e.g. urban modelers, hydrologists, biologists). This project will support an interdisciplinary framework to analyze and predict land use dynamics. It will also enhance geographical knowledge and theories, as well as applications to societal issues, such as urban sprawl, environmental degradation and sustainable development.