Extracting Impervious Surface from Aerial Imagery Using Semi-Automatic Sampling and Spectral Stability

Hua Zhang,Steven M Gorelick,Paul V Zimba

doi:10.3390/rs12030506

Hua Zhang, Steven M Gorelick + Show 1 more

Open Access

https://doi.org/10.3390/rs12030506

Copy DOI

Abstract

The quantification of impervious surface through remote sensing provides critical information for urban planning and environmental management. The acquisition of quality reference data and the selection of effective predictor variables are two factors that contribute to the low accuracies of impervious surface in urban remote sensing. A hybrid method was developed to improve the extraction of impervious surface from high-resolution aerial imagery. This method integrates ancillary datasets from OpenStreetMap, National Wetland Inventory, and National Cropland Data to generate training and validation samples in a semi-automatic manner, significantly reducing the effort of visual interpretation and manual labeling. Satellite-derived surface reflectance stability is incorporated to improve the separation of impervious surface from other land cover classes. This method was applied to 1-m National Agriculture Imagery Program (NAIP) imagery of three sites with different levels of land development and data availability. Results indicate improved extractions of impervious surface with user’s accuracies ranging from 69% to 90% and producer’s accuracies from 88% to 95%. The results were compared to the 30-m percent impervious surface data of the National Land Cover Database, demonstrating the potential of this method to validate and complement satellite-derived medium-resolution datasets of urban land cover and land use.

Highlights

Over half the Earth’s population resides in cities [1]
The resolution mismatch between 1-m National Agriculture Imagery Program (NAIP) imagery and 30-Landsat imagery was evident in such areas and had an apparent impact on the spectral characteristics of training samples
A hybrid remote sensing method was developed to improve the extraction of impervious surface from aerial imagery

Summary

Introduction

Over half the Earth’s population resides in cities [1]. Urbanization inevitably alters hydrologic processes, energy balance, and biological composition, resulting in higher nutrient loads, elevated surface temperature, increased peak flow, and accelerated habitat degradation in many urban areas and watersheds [2,3,4]. This has led to the generation of important databases at national or global scales, such as the National Land Cover Database (NLCD) in the United States [12,13], Coordination of Information on the Environment (CORNIE) land cover inventory in Europe [14], Finer Resolution Observation and Monitoring-Global Land Cover (FROM-GLC) database in China [15], and MODIS Land Cover Type Yearly Global dataset [16] As these databases are derived from medium-resolution satellite imagery with pixel size ranging from 30 m to 500 m, they have to use multiple urban land cover classes (i.e., low- and high-intensity developed areas) to reflect different levels of impervious surface cover at the scale of a satellite image pixel. Random forest (RF) models were combined with geographic object-based image analysis to generate a 1-m land cover dataset of West Virginia with an overall accuracy of 96.7% [19]

Objectives

Methods

Results

Discussion

Conclusion