Authors: Karim Malik*, , Colin Robertson, Wilfrid Laurier University
Topics: Spatial Analysis & Modeling, Geographic Information Science and Systems
Keywords: Landscape similarity, pattern detection, fine-grain texture features, texture-based convolutional neural network
Session Type: Paper
Start / End Time: 5:00 PM / 6:40 PM
Room: Washington 6, Marriott, Exhibition Level
Presentation File: No File Uploaded
Convolutional neural networks (CNN) are known for their ability to learn shape and texture descriptors useful for object detection, pattern recognition and classification. The deeper layer features of a CNN learn global image information vital for whole scene or object discrimination. In landscape pattern comparison, however, dense localized information encoded in shallow layers can contain crucial information for characterizing changes across images local regions, but are often lost in fully connected layers. In this paper, a texture-based CNN is constructed via a feature concatenation and pooling framework which results in capturing dense texture features. The traditional CNN architecture is adopted as a baseline for comparison. Experiments of our model architecture are demonstrated with both simulated texture data at geographic scales, and very high resolution unmanned aerial vehicle imagery. A sensitivity analysis to variable architectures of the networks and model parameterization is carried out. Visual examination of the patterns learned by the CNN filters using gradient-based class activation maps reveals that the features conform meaningfully with patterns found in training samples, but the texture-encoded CNN features maybe more discriminative and useful for similarity mapping as they are largely fine-grained. Further research into learning specific distinctive geographic texture types to test the effectiveness of network layers in yielding features relevant to pattern detection is needed. We conclude with a discussion of implementation issues in a landscape monitoring context.