Optimizing Deep Learning Model Selection for Angular Feature Extraction in Satellite Imagery

摘要

Deep learning techniques have been leveraged in numerous applications and across different data modalities overthe past few decades, more recently in the domain of remotely sensed imagery. Given the complexity and depthof Convolutional Neural Networks (CNNs) architectures, it is difficult to fully evaluate performance, optimize thehyperparameters, and provide robust solutions to a specific machine learning problem that can be easily extendedto similar problems, e.g. via transfer learning. Ursa Space Systems Inc. (Ursa) develops novel machine learningapproaches to build custom solutions and extract answers from Synthetic Aperture Radar (SAR) satellite datafused with other remote sensing datasets. One application is identifying the orientation with respect to truenorth of the inlet pipe, which is one common feature located on the top of a cylindrical oil storage tank. In thispaper, we propose a two-phase approach for determining this orientation: first an optimized CNN is used toprobabilistically determine a coarse orientation of the inlet pipe, followed by a maximum likelihood voting schemeto automatically extract the location of the angular feature within 7:5°. We present a systematic technique todetermine the best deep learning CNN architecture for our specific problem and under user-defined optimizationand accuracy constraints, by optimizing model hyperparameters (number of layers, size of the input image,and dataset preprocessing) using a manual and grid search approach. The use of this systematic approach forhyperparameter optimization yields increased accuracy for our angular feature extraction algorithm from 86%to 94% and can be extended to similar applications.