Courtesy Photo | Deep learning model approach. Image of volcanic plume is fed to the machine learning...... read more read more
Subscribe
Machine learning is cutting through the fog in aerial dispersion of chemical threats.
Understanding and predicting the dispersion and concentration of chemical warfare agents is a central factor in protecting the Joint Force. Currently, military commanders and staff plan large-scale combat operations using a best-guess method of where the enemy would place their chemical weapons on the battlefield. Reconnaissance teams in protective gear would confirm these estimates.
To improve the odds of locating chemical weapons, the Defense Threat Reduction Agency’s (DTRA) Chemical and Biological Technologies Department in its role as the Joint Science and Technology Office (JSTO) for Chemical and Biological Defense, an integral component of the Chemical and Biological Defense Program, partnered with scientists at the Pacific Northwest National Laboratory (PNNL) to uncover the capabilities of deep learning, a subset of machine learning, in the domain of hazardous material quantification using 2D images, such as photographs, of chemical plumes.
Although not always, deep learning is often formed through the implementation of a neural network, which is a system of mathematical layers that identify and extract important features from nearly any kind of data, such as numerical, categorical, image, audio, and video. A Convolutional Neural Network (CNN) interprets individual pixels from images and assigns weights to specific features or groups of pixels in each image to assist in classifying the entire image. The CNN model is first trained on a set of labeled images. This means that the model is provided with the correct category for each of the initial images. Once the model is sufficiently trained, it can identify similarities and key features in the dataset and can then calculate the probability that the image belongs to any number of existing categories. The category with the highest resulting probability is the one assigned to the image.
In collaboration with JSTO, PNNL researchers created a CNN model designed to classify plumes so it could characterize chemical plumes. They used images pulled from publicly available sources that simulated the plumes exhibited by many chemical warfare agents. The collected photos are a combination of ground, aerial, and satellite imagery of multiple volcanic plume events. The results of this research preliminarily demonstrated a high efficacy of deep-learning models in classifying volcanic plume data, which can reduce human biases and errors.
JSTO plans to apply the final CNN model to the problem of quantifying hazardous materials on the battlefield from two-dimensional images. This ability will enable commanders in a post-engagement environment to better assess follow-on maneuvers to counter the adversary’s use of chemical area-denial systems in persistent (long lasting) or nonpersistent (short-term) forms. Through this technique, commanders can preserve combat power and maximize capabilities against continued use of chemical weapons against the U.S. Joint Force and allied forces. The results from this study could lead to further integration of deep learning into military decision-making processes and create the opportunity for Joint Force leaders to plan military operations with more accurate chemical threat predictions.
Machine learning is an emerging discipline that ensures more efficient and more accurate analyses of complex topics. JSTO’s exploration of machine learning capabilities in countering chemical weapons can enable more reliable modeling and forecasting of threats to warfighters on the battlefield and better-informed decisions for Joint Force commanders.
POC: Richard Fry, richard.n.fry.civ@mail.mil
| Date Taken: | 04.24.2024 |
| Date Posted: | 04.24.2024 16:51 |
| Story ID: | 469441 |
| Location: | FT. BELVOIR, VIRGINIA, US |
| Web Views: | 427 |
| Downloads: | 0 |
This work, Finding Data in the Drift, must comply with the restrictions shown on https://www.dvidshub.net/about/copyright.
No keywords found.
