Identifying Smokestacks in Remotely Sensed Imagery via Deep Learning Algorithms

Locating smokestacks in remote sensing imagery is a crucial first step to calculating smokestack heights, which allows for the accurate modeling of dioxin pollution spread and the study of resulting health impacts. In the interest of automating this process, this thesis examines deep learning networks and how changes in input datasets and network architecture affect image detection accuracy. This initial image detection serves as the first step in automated object recognition and height calculation. While this is applicable to general land use classification, this study specifically addresses detecting smokestack images. Different dataset scenarios are generated from the massive Functional Map of the World dataset, ranging from two to sixty-two classes, and network architectures from recent studies are used. Each dataset and network is analyzed in their performance by way of F-measure. Image characteristics are also analyzed from images that were correctly/incorrectly labeled by the algorithms, providing answers on what images the algorithms best predict and what qualities the algorithms cannot discern. The smokestack’s accuracy is reported at its highest through a five class training dataset, using an Adam Optimizer over six epochs. More or less classes returned lower scores, as did using the Stochastic Gradient Descent optimizer. Extended epochs did not return significantly higher or lower scores. The study concludes that while using more data can be effective in creating more accurate algorithms, using less data which is better structured for the problem at hand can have a greater effect.