Reinforcement Learning Based Optimal Adversarial Pathway Estimation Using Remotely Sensed Spectral-Terrain Data and Human Value Assessment

Abstract

Extended Abstract Geo-intelligence organizations are often faced with the need to determine optimal pathways that adversaries may take based on various types of information including remotely sensed imagery and human geo-intelligence. The mobile enemy problem, where the objective is to predict the pathway that a mobile enemy may take, is considered here as a way to develop a statistical/signal processing formulism to assist leadership in making better decisions about how to estimate the whereabouts of an adversary. A two-tier processing pipeline utilizing feature extraction and reinforcement learning-based optimal pathway estimation was created to demonstrate how human/machine learning teaming can be exploited to address a geo-intelligence problem. The information used in the processor development consists of an open-source hyperspectral imagery (HSI) data set [1]. A strip map of terrain HSI was divided into 32 x 32 pixel image chips where principal component analysis [2] was used to reduce the dimension and decrease the noise of the hyperspectral signatures. Spectral dictionary endmembers [3] were estimated from the denoised HSI data using the unsupervised learning algorithms of k-means clustering [4] and automatic target generator processing [3]. This substage was necessary in order to perform image chip value estimation. In this evaluation stage, five different algorithms were used to calculate different value fields. Each technique used a feature extraction method designating the relative value of each image chip comprising the complete HSI scene. The first algorithm used for HSI image chip value estimation consisted of abundance estimation via nonnegative constrained least squares matched filtration [3] along with a