How Machine Learning is Changing Intelligence Collection

There is an extraordinary amount of data being generated around the world on a daily basis. The task of collecting relevant information, organizing it, and piecing it together in a way that tells a story seems like an overwhelming and nearly impossible task. Yet, this is the monumental task of intelligence analysts.

Thankfully, incredible technological advancements, including machine learning (ML) and artificial intelligence (AI), are assisting analysts in their efforts to collect and categorize massive amounts of data. Technology is evolving at a rapid pace, and analysts must always be learning how to apply machine-learning technology to help them better understand and solve complex problems.

When I was a student earning a master’s degree in Intelligence Studies at American Military University (AMU), I learned about these technologies and how critical they were for the intelligence profession. As an intelligence professional today, I was able to take what I learned in the classroom and apply it directly to my job.

One of my primary job responsibilities is to collect many types of data such as spatial data, which also includes satellite imagery. I use these data sources to look for objects of interest. However, it can take an analyst a very long time to visually and manually scan individual images for specific objects. Using ML and AI can automate object detection and enhance the efficiency at which intelligence is derived.

In order to use these technologies, an analyst must first start by “training” data, which includes a step called “labeling.” In this step, analysts have pre-captured image examples about what a particular object looks like.

For example, say the objects of interest are buildings. The analyst finds several example images containing building types. The first step is to capture the many different attributes of a building and add it into the labeling software (see image below). Such attributes include types of buildings such as residential, military, or commercial. In the training images, the analyst finds buildings and “annotates” them using rectangles or polylines, and then chooses the classification (type) of the object based on the attributes. The image is now labeled and included in the software.

Characteristics of features are captured and placed in the left column for analysts to attribute features. (Image used with permission by LabelBox).

This step is repeated over and over for as many images are needed (usually in the hundreds or thousands). The more training data input into the system, the better the output results will be. These “labels” are then exported as a JavaScript Object Notation (JSON) file, which is then applied to the development of a machine learning model. These models are then used to infer objects within images—a term referred to as inferencing.

Then, an analyst can simply connect a database of images that haven’t been reviewed yet and the computer will scan each image, detect the object (in this case, buildings), auto-classify them, and return the found objects (flagged in the image) to an analyst for review. Using this technology can provide immense enhancements in processes, workflows, and quality output.

Even with this technology at the fingertips of an analyst, it important to remember that intelligence remains an art.