Semantic segmentation of aerial imagery using color-texture analysis and convolutional neural networks

The wide availability of large volumes of high resolution remotely sensed data have increased a number of research avenues and data driven applications in the fields of image processing and computer vision. Specifically, pixel-wise image segmentation is a challenging task as each pixel is assigned a label based on a local context surrounding the pixel. Texture plays a critical role in image segmentation as it is one of the most important visual cues in identifying homogeneous areas. This is especially true in aerial imagery where environmental characteristics such as vegetation, water, pavement, are often non-uniform and have unique texture composition due to changes in orientation, scale, or visual appearance.

Such data play a pivotal role for the Military and contribute to overall situational awareness. This work presents the use of discrete wavelet decomposition with a fully convolutional neural network to address the need for an automated pixel-wise image segmentation approach for aerial imagery collected by UAS's. Specifically, it highlights the implementation of wavelet decomposition as a pre-processing step with a pretrained CNN algorithm and use of transfer learning to train a deep neural network for pixel classification in RGB imagery.