Atmospheric Correction Algorithms for Improved Settlement Mapping and Building Detection Performance

Jeanette Weaver

Jeanette Weaver Oak Ridge National Laboratory
Hsiuhan Lexie Yang Oak Ridge National Laboratory
Katie Corcoran Oak Ridge National Laboratory
Jacob McKee Oak Ridge National Laboratory

14D

With the significant advances in imaging technologies, high resolution multi-spectral imagery has become invaluable in understanding our environment. In a passive sensing scenario, the complex path between energy sources to a sensor detector and atmospheric variations create the need for quality atmospheric correction within an efficient pre-processing workflow. For broad scale monitoring or land use classification, images covering the region of interest are usually acquired at different dates due to the swath pattern and revisit rate of a particular sensor. With these multi-temporal images, utilizing remote sensing data for these applications requires careful calibration of pixel values to consistent sensor measurements, particularly for the purpose of mitigating affects of seasonality and atmospheric variability. In this research we specifically investigate the importance of well-calibrated images in two large scale image analysis tasks: human settlement mapping and building detection.
Previous research has been proposed to address the issue of multi-temporal image analysis. However, the temporal images were most often exploited as the form of raw digital counts, posing certain limitations when distribution shift is significant, particularly in a semi-supervised learning setting where learned classifiers (or models) are dynamically altered to fit the new image. This issue is even difficult to overcome when a large collection of images covering extended area exhibits great spectral variability. We argue that implementing proper atmospheric corrections, such as IARR and Log Residuals in a pre-processing workflow would increase the performance of machine learning designs for settlement and building extraction. The settlement and building extraction methods are built on a Support Vector Machine semi-supervised classification and a Deep Learning CNN approach respectively. Four methods are tested and reviewed, including: IARR, Log Residuals, FLAASH, and QUAC. These methods are evaluated based on the resulting accuracy from settlement and building extraction.

Keywords: high resolution multi-spectral imagery, atmospheric correction, deep learning, support vector machine, building detection

Remote Sensing Division
Primary Data Acquisition Division

14:30 Atmospheric Correction Algorithms for Improved Settlement Mapping and Building Detection Performance, Jeanette Weaver

January 29 @ 14:30
14:30 — 14:45 (15′)

Mineral A

Jeanette Weaver

Add to Google Calendar