Quantum Boosting and Fast Classical Metrics for Tree Cover Detection in Remote Sensing Data

Monday, 15 December 2014: 3:25 PM
Edward Boyda1, Saikat Basu2, Sangram Ganguly3, Andrew Michaelis4 and Ramakrishna R Nemani3, (1)Saint Mary's College of California, Moraga, CA, United States, (2)Louisiana State University, Computer Science, Baton Rouge, LA, United States, (3)NASA Ames Research Center, Moffett Field, CA, United States, (4)University Corporation at Monterey Bay, Seaside, CA, United States
New volumes of high resolution remote sensing imagery hold great
promise for Earth science, and with it, new challenges in machine
learning. Familiar heuristic training routines become impractical as
datasets scale to terabytes and beyond. Now, emerging quantum
hardware from D-wave Systems allows us to explore alternatives based
on the principles of adiabatic quantum computation. As part of a
program to develop tree cover estimates for the continental United
States based on one-meter-resolution National Agriculture Imagery
Program (NAIP) data, we have implemented a binary classifier, known as
Qboost, to combine in a principled manner decision stumps defined
on features extracted from 8x8 pixel squares. Qboost was originally
developed to be trained on D-wave hardware. Prototyped on NAIP data
for the state of California, the classifier discrimates tree-covered regions with a validation
error rate of 8%. Additionally, we identify quadratic combinations
of the Atmospherically Resistant Vegetation Index (ARVI) and standard
deviations of intensity or near-infrared reflectance that provide
fast, simple, classical metrics to identify tree cover. They cut by nearly half the
error rates of ARVI used alone or of our best single-feature