Title: Spectroscopic Image Signature Classification of Land Cover Types using Multi-Spectral Data within a Neural Network

Keywords: Artificial neural networks, Hyperspectral imaging, Spectroscopy, Feedforward neural networks, Multispectral perceptrons, Neural networks, Uniform resource locators, Random Access Memory, Software

Abstract:
Through improvements in technology, high resolution multi-spectral imaging allowed new capabilities to become available in the remote sensing field.  Spectral signature classification technologies existed in the chemical spectroscopy field to identify minerals by way of active systems [1].  The theory of this paper surrounded the premise that passive systems can provide spectral signatures of objects within images from satellite platforms. Specifically this paper targeted land cover types from the Kittyhawk, North Carolina area. Multi-spectral signals presented up to seven individual readings per pixel. As the decision support system, a neural network was trained to decide the type of land cover based on the band readings. In an effort to determine specific land cover types based on need, ground truthed spectral readings were also classified using a linear model to convert the readings into approximate satellite readings. The converted readings were then classified by the trained neural network. A minimal r-squared valued of 86% was required to be considered a viable method of image classification.

The methodology was carried out as followed at the first field sample location was considered pixel one at zero meters.  After that, two more sample points were taken from that pixel at thirty-two meters and sixty-four meters linearly.  At the last point in the pixel, one-hundred thirty-two feet was measured from that pixel to the next pixel, and the process repeated for the collection of the remaining sample points.

Spectral readings were taken from two locations in North Carolina, including the Sand Dunes of Jockey’s Ridge and the fields of the Wright Brothers’ Memorial Park.  In order to have access of the public land areas, research permits were requested, completed, and presented to the respective park rangers at each location {Appendix A and B}.  The spectral readings were saved as tab delimited files with an integration time of 50ms within the “spect” sub-folder of the pixel folder within the file system. The coordinates were marked on the GPS device in decimal format and later transferred to Google Earth. The reference image of the sample location was lastly taken.  The images were saved with a timestamp within the “pict” sub-folder of the pixel folder within the file system.

These GPS coordinates imported into Google Earth from the GPS device were saved in keyhole markup language (KML) file. The KML file was then entered into an application named KMLCSV Converter, to export the decimal values of the GPS coordinates into a comma separated values (CSV) format. Landsat GeoTIFF imagery was retrieved from the USGS Global Visualization Viewer (GLOVIS) website.  These Landsat images were then opened in ENVI to gather the pixel band data. In this specific case, four different bands of the possible seven readings were recorded for each pixel. This limitation occurred because the lab spectrometer’s upper wavelength limit was 65535 illuminant readings. The data was then entered into the master Google Spreadsheet document.How to develop a sensor platform housing lab spectral equipment enabling use for field work?  An instrumentation platform named “The Spectator was constructed in order to house the lab spectral equipment for field work.
What would be the best possible workflow to collect data using available equipment?  The best possible workflow for data collection was to save the spectrum as tab delimited files, take time and date stamped reference images, mark the GPS coordinates on the Garmin GPS systems, import those coordinates into Google Earth and convert those save files using the KMLCSV Converter Software, and use Exelis ENVI 5.0 to gather the Landsat band data utilizing the GPS coordinates.  Lastly, Google Spreadsheet was utilized to enter in the data, and Microsoft Excel was utilized to perform a linear regression on the datasets to convert the Spectrometer Illuminants Values to Landsat Brightness Values.
What correlation exists between Landsat readings and spectrometer readings if any?  The correlation that exists between Landsat readings and spectrometer readings was that the Landsat Brightness values, ranging from 0-255, correlated with the Spectrometer Band Illuminants Values, ranging from 0-65535.
How can a neural network be utilized to perform land cover classifications?  An arff file was created in order to feed the Landsat band data and the land cover classification types into the Multilayer Perceptron Function to classify the land cover types.
Can a neural network classify land cover types with at least 86% accuracy?  Yes, a neural network can classify land cover types with at least 86% accuracy.

Reccomendations for future works would include development of a larger data set sample with varied land cover types is needed in order to exhaustively investigate the classification of land cover types using spectral signatures.  Specifically mixed pixel classifications should be explored using the spectral signature technique.  The lab spectrometer used limited the data, allowing sample readings to reach up to only four out of the seven bands that existed. Expanded spectral equipment should be obtained in order to carry out the investigation.  Having access to X-band (hyper-spectral) high resolution data or writing an image grant far in advance of the investigation to achieve access to a high-resolution satellite in which to verify the spectral readings collected from the ground truth data is necessary for achieving truly unique spectral signatures.  Development of a software macro or all inclusive application to cut down on data collection time is highly recommended. Developing a program that would take a picture and name that picture; take the spectral signature and name that spectral signature; and mark the GPS coordinate, name the GPS coordinate, and save the GPS coordinate as a KML (keyhole markup language) file would be invaluable to future researchers. Using a netbook or a full blown notebook with higher system specifications and minimal programs installed on the device is also highly recommended allowing for smooth and quick data collection.