A geographic computational visual feature database for natural and anthropogenic phenomena analysis from multi-resolution remote sensing imagery

Abstract

Remote sensing imagery and sensor data comes in a variety of spatial and spectral characteristics, including very high resolution images from spaceborne and airborne sensors, as well as high, medium, and low resolution sensors from space. In image information mining, the resolution of the imagery affects the characteristics of the features, information, and knowledge that can be extracted or harvested from the imagery pixels. For instance, low resolution imagery has salient visual features on the scale of urban areas, rivers, coastlines, and other geographical land forms. These large-scale features are great for environmental monitoring and understanding large-scale geomorphic processes. However, high resolution imagery allows application tasks such as the extraction of building footprints or counting vehicles in a parking lot. For this reason, different resolutions of imagery and different image processing techniques are often necessary. To develop a truly flexible and robust modeling capability, it is necessary to mine image features and information across a variety of scales and using a variety of image processing and computer vision techniques, each appropriately attuned to the type of information and the spatial-spectral characteristics of the imagery. In this research, we develop a multi-resolution, multi-sensor framework for processing a variety of remote sensing imagery, including MODIS, Landsat, and high-resolution satellite imagery. Deep neural networks are used to extract image features from high-resolution images; while classical image feature extraction techniques are applied to medium and low resolution sensor data. The various features are agglomerated into a geospatial data cube using PostGIS, with the aim of facilitating advanced geospatial modeling of natural and anthropogenic processes. We demonstrate how this multi-resolution data cube of remote sensing visual features facilitates analysis of natural and anthropogenic phenomena, as well as discuss some potential future applications.