The Tradecraft of Artificial Intelligence and Machine Learning
The integration of AI and machine learning into GEOINT tradecraft presents a significant paradigm shift
By Mike Rampino, Preferred Systems Solutions; Steven R. Thomas, Ball Aerospace; Stephen H. Tupper, Missouri University of Science and Technology; Marion Neumann, Washington University in St. Louis; and Peter Morosoff, Electronic Mapping Systems, Inc. (E-MAPS)
The tradecraft of geospatial intelligence (GEOINT) is always evolving. However, the integration of artificial intelligence (AI) and machine learning (ML) into GEOINT tradecraft presents a significant paradigm shift, and like previous technical innovations that dramatically change and advance the tradecraft, a thoughtful, broad-reaching approach to the adoption of these technologies is necessary. AI and ML go beyond the introduction of technical innovation such as the conversion of film and print media to digital media or 3D stereoscopic capabilities.
The introduction of AI and ML into GEOINT will cause analysts and practitioners to interact with technology in a new way. In addition to learning new technical skills they will learn to teach geospatial science to AI. They will also oversee geospatial workflows and practices to determine where AI and ML can be inserted into processes to provide automation and augmentation. The merger of AI and ML within the GEOINT tradecraft will continue to advance toward a place in which its practitioners possess the knowledge and skills to be a steward of the GEOINT practice and the practitioner can leverage AI and ML to create new points of innovation. In the early stages of this inclusion of AI and ML we can already identify strong steps being made where Data Scientists work alongside GEOINT analysts to achieve mission outcomes.
- This article is part of USGIF’s 2019 State & Future of GEOINT Report. Download the PDF to view the report in its entirety.
The defense and intelligence communities have previously described enhancements of system performance and functionality in existing or deployed capabilities by inserting new or significantly improved technology. A vertical insertion enhances a single capability from bottom to top at components, equipment, subsystems, systems, system of systems, and kits. A horizontal insertion is the utilization of a new or improved technology in similar but distinct platforms or disciplines. The GEOINT Community should view the incorporation of AI and ML as the latter.
Historically, horizontal insertion of new technology can require a full generation to achieve. This is caused by an insertion model that waits for senior personnel to retire and entry-level personnel are the focus of training on the new technology. The GEOINT Community does not have a full generation to incorporate AI/ML technology. Insertion of AI/ML within the GEOINT tradecraft must move faster to keep pace with the exponential growth of data collected and to stay a step ahead of U.S. adversaries. If the GEOINT Community waits a generation to fully incorporate AI/ML, we will become irrelevant (and perhaps be dominated by our adversaries). Thus, new and aggressive education and insertion models must be adopted.
Recent history provides many examples of new technologies being adopted for national security purposes. Often, complex scientific and engineering concepts have been translated into layman’s terms to enable training forces to employ new weapons or new enabling capabilities. For example, maritime navigation is based on geophysics and other scientific principles that might require an advanced degree to fully comprehend. Yet, the National Geospatial-Intelligence Agency (NGA) and its predecessor organizations have for years produced a widely used reference for laymen without such advanced degrees who successfully navigate the world’s oceans. The adoption and operational employment of RADAR in World War II and the operational deployment of nuclear weapons after World War II provide other examples. In each case, doctrine, training, and procedures had to be developed and implemented to allow airmen, sailors, marines, and soldiers with relatively little scientific or engineering knowledge to successfully operate complex and potentially lethal systems.
To be successful, the GEOINT Community must create a culture within the tradecraft in which analysts and practitioners come to trust automated systems. It must cultivate a culture that has an eagerness to use AI/ML to replace manual, human-driven processes. The GEOINT Community must grow beyond its current educational programs and credentials to include new skills and knowledge. It must integrate the skills that support AI/ML within existing education and training programs. To achieve accelerated adoption of AI/ML, the GEOINT Community requires a multi-echelon educational offering related to AI/ML technology.
These echelons are nested such that tradecraft practitioners at various seniority levels and of varying types of expertise receive tailored education and training that provide them the skills to employ AI/ML approaches such as using database platforms, structuring data warehouse environments, information storage and retrieval systems, web search engines, text mining, collaborative filtering, and recommender systems. These entry-level tasks may be appropriate subjects for instruction at the associate degree-level or in the form of industry certifications focused on specific hardware and software. These base-level skills in both hardware and software have a shorter shelf life due to constant improvement and rapid expansion.
At the next level up are the data scientists. They are likely to need a mix of bachelor’s and master’s degree-level understanding of regression, classification, resampling methods, model selection, regularization, decision trees, support vector machines, principal component analysis, and clustering. Analysts who draw on data science talent must first know the GEOINT domain and will succeed through collaboration with data science models and tools. GEOINT analysts in collaboration with data scientists will need to draw upon their combined talents and expertise to operate AI/ML comfortably across the GEOINT mission.
Beyond analysts, the top-echelon of decision-makers will require special instruction and education. Executives are drawn from many disciplines and don’t necessarily lead the ranks they grew up in. It is more likely they have a variety of experiences in many fields and will have to be coached, more than educated, in how to best understand AI/ML-derived interpretations. Here the transition state equals the end state. High-level decision-makers are to be helped by learning an overarching understanding of the tradeoffs of using AI and ML, understanding the nuance associated in accepting AI/ML-augmented processes and products, and being prepared to invest in the maturation of the art and science of interpreting data via machines.
At the outset of using AI/ML within GEOINT processes, analysts, engineers, supervisors, and executives all need to understand that a product or recommendation for decisions based on AI/ML-dependent analysis should be treated with caution, possibly needing more verification by experienced humans until a consistent record of prediction has been statistically correlated with established tradecraft techniques. At the same time, these practitioners must be given training that allows them a depth of understanding that supports a willingness to invest in refining processes, algorithm development, datasets, etc. Additionally, this education needs to provide the fundamental acumen on which they can measure the maturity of the inserted AI/ML technology.
At another scale, an analyst should have a very different training in the AI/ML system—perhaps how it is coded, or the selection of filters, the segmentation of data, the speed of analysis, and the comparison of error. Within the GEOINT Community each practitioner (i.e., manager, engineer, data scientist, and analyst) must work together, leveraging their different skills and expertise to improve the technology through methodologies such as mining, scraping, manipulating, transforming, cleaning, visualizing, summarizing, and modeling large-scale data as well as supervised and unsupervised machine learning algorithms applied in various mission scenarios.
AI and ML have the potential to greatly improve the productivity, capacity, and capability of GEOINT analysts, enabling them and their organizations to capitalize on the ever-increasing amount of data available. In the near-term, advances in computational power, artificial neural networks (ANNs), and computer vision enable new approaches to GEOINT tradecraft. NGA Director Robert Cardillo has said eight million more GEOINT analysts would be needed to analyze all the data expected to be available as remote sensing systems and other geospatial data sources proliferate. Since educating, training, and employing millions of additional GEOINT analysts is unlikely if not impossible, incorporating AI and ML into GEOINT tradecraft might help us keep up. But discussions of how to best incorporate AL and ML into GEOINT tradecraft can reveal disparate views.
Some assert that anyone wishing to apply AI/ML must have an advanced degree in computer science, math, or statistics and be proficient in coding and writing software. The thinking is it would be dangerous for anyone without such education and skills to apply AI and ML. Such an approach would certainly provide practitioners greater confidence in applying AI/ML to GEOINT tradecraft, but it would likely also significantly slow speed of adoption. We might also find that people eager to be GEOINT analysts don’t necessarily have the same passion for being computer scientists or mathematicians.
In order to successfully determine where AI and ML can be inserted into GEOINT processes, engineers and practitioners tasked with its implementation or development need to gain a substantial understanding of the fundamentals of AI/ML algorithms. This typically requires a solid background in probability and statistics, linear algebra, and calculus. Proficiency in probability and statistics is not only important for engineers who want to understand and implement AI/ML methods, but it is also a critical skill for analysts and end users who apply AI/ML methods—even if the methods themselves are treated as a black box. Users of AI/ML techniques need to understand, interpret, and judge both input and output to AI/ML algorithms applied to practical problems.
The educational echelons of the GEOINT Community will need to ensure fundamentals such as linear algebra and calculus, which are foundational to the understanding of AI/ML algorithms. Conversely, the developers and engineers tasked with implementation of AI/ML technology, whether from scratch or existing implementations, are approaching AI/ML from a computer science perspective. They require proficiency in data structures and algorithms (including complexity analysis).
As there is no one ML method that solves all problems, engineers will have to acquire a basic understanding of the strengths and weaknesses of the state-of-the art methods. Further, it is important to understand the ML workflow and how to evaluate and compare algorithms in a sound and scientific manner as well as how to internalize the process of comparing and evaluating algorithms on various application domains. Engineers will have to dive deeper into the learning algorithms that typically leverage non-linear optimization and advanced calculus. At the core is a focus on understanding, implementing, and analyzing AI/ML algorithms, however related fields of study such as computer vision, big data processing, and cloud computing should be considered in a holistic AI/ML education.
By recognizing the different needs of GEOINT Community, a multi-echelon educational approach advocates teaching AI/ML as a series of courses or programs that allow students to achieve the level of familiarity with AI/ML methods their role within the GEOINT Community requires. Providing multiple courses, paths, and tracks covering the introduction of AI/ML at undergraduate and graduate levels ensures the variety of roles, positions, and seniority levels within the community are provided the education and training needed to successfully adopt AI/ML.
Growing Confidence in AI/ML
The community is in the early phase of applying AI/ML to GEOINT tradecraft. Defense and intelligence organizations such as NGA have pilots underway that should shed light on the best approaches. These pilot programs have helped reveal and identify challenges in inserting AI/ML into GEOINT workflows. Some of these challenges include but are certainly not limited to data scarcity, lack of data diversity, difficulty in scaling AI/ML, and legacy systems that were designed around human perception and performance. Each of these challenges must be overcome to fully realize the benefits of AI/ML.
However, perhaps the greatest challenge from the perceptive of the GEOINT tradecraft is that of confidence in use of the emerging technologies. AI/ML offers a future in which analysts are freed from much if not all of the manual data management tasks that consume a large amount of their time. They are freed from tasks such as data labeling and allowed to focus on mission-related analysis and production. However, those analysts must have confidence in the AI system.
In these early days of applying AI/ML to GEOINT tradecraft, it seems teaming analysts with data scientists is yielding successes. The GEOINT analysts have seen significantly increased productivity and are confident in applying ML to their analytical problems. Today, GEOINT analysts participating in these pilot programs depend on close collaboration with data scientists. The data scientists develop models and implement ML algorithms. GEOINT analysts work with the data scientists to help validate the models but the data scientists do the development and write the code. The collaboration seems to be instilling a level of understanding and confidence in AI/ML. In the longer-term, when AI/ML tools and processes are implemented at an enterprise scale, the GEOINT Community will need to determine how to build confidence in its analysts and leadership and determine whether constant collaboration with data scientists will diminish over time or become an institutionalized change within the community’s tradecraft.
 “The American Practical Navigator” was first published in 1802 and was most recently published by NGA in 2017. It had been published by NGA predecessor organizations for decades. NGA provides access to the material on line at https://msi.nga.mil/NGAPortal/MSI.portal?_nfpb=true&_pageLabel=msi_portal_page_62&pubCode=0002.
 The United States Air Force initially used highly-educated contractors to staff its first Atlas Intercontinental Ballistic Missile (ICBM) system but then transitioned to less-technically sophisticated and educated military operators after a Strategic Air Command crew completed a successful training launch. Jacob Neufeld, “Ballistic Missiles,” (Washington, D.C., Office of Air Force History, 1990), 103, 208, 252-253.
 Robert Cardillo, Director of the National Geospatial-Intelligence Agency, remarks delivered at the GEOINT 2018 Symposium, 23 April 2018, available at https://www.nga.mil/MediaRoom/SpeechesRemarks/Pages/GEOINT-2018-Symposium-.aspx.
Do We Need A “Civilian ARPA” for AI?
The case for leveraging artificial intelligence to improve public service
Making Sense of Compression
Understanding the three common types of raster image compression