Improved and More Robust Automatic Target Classifiers

Navy SBIR 20.2 - Topic N202-120

Naval Air Systems Command (NAVAIR) - Ms. Donna Attick [email protected]

Opens: June 3, 2020 - Closes: July 2, 2020 (12:00 pm ET)



N202-120       TITLE: Improved and More Robust Automatic Target Classifiers


RT&L FOCUS AREA(S): Artificial Intelligence/ Machine Learning, General Warfighting Requirements (GWR)

TECHNOLOGY AREA(S): Air Platform, Information Systems, Battlespace


The technology within this topic is restricted under the International Traffic in Arms Regulation (ITAR), 22 CFR Parts 120-130, which controls the export and import of defense-related material and services, including export of sensitive technical data, or the Export Administration Regulation (EAR), 15 CFR Parts 730-774, which controls dual use items. Offerors must disclose any proposed use of foreign nationals (FNs), their country(ies) of origin, the type of visa or work permit possessed, and the statement of work (SOW) tasks intended for accomplishment by the FN(s) in accordance with section 3.5 of the Announcement. Offerors are advised foreign nationals proposed to perform on this topic may be restricted due to the technical data under US Export Control Laws.


OBJECTIVE: Develop better and more robust automatic target classifiers capable of providing improved accuracy, identification, and classification of complex or subtle dynamics by leveraging advanced mathematical and machine learning tools.


DESCRIPTION: Current tactical platforms are challenged when it comes to target identification and classification algorithm development. They are unlikely to routinely encounter more complex dynamics of targets of interest and when they do, the raw data is not likely to be recorded. Therefore, data from other collection systems and/or computer models must be used to model and simulate the dynamics and build the required algorithms. The advancement of powerful super computers has made near-real physical modeling possible [Ref. 1], allowing modeling of almost any target with its environment and achieving very good agreement between models and observations. It is important to note though, some approximations are usually required but those terms are generally small and are usually considered insignificant.


Advanced mathematical and machine learning techniques may be used to resolve this apparent paradox between exploiting a high-dimensional feature space with data intensive machine learning and a lack of understanding of the underlying dynamics. In addition, these techniques can enable classification somewhere near real-time, a timescale that is relevant to tactical platforms, that is on the order of seconds. With this approach, one could build and train equivalently effective algorithms with built-in physics, i.e., coupled non-linear differential expressions, to ensure the algorithms are robust. Finally, the learned physics-based models could be used to extend accurate classification to other objects of similar class using sparsely sampled data, computer models, and scaled model data.


Machine learning techniques, e.g., Support Vector Machines (SVM), Dynamic Mode Decomposition (DMD) [Ref. 5], Long Short-Term Memory (LSTM) Recurrent Neural Networks (RNN) or Convolutional Neural Networks (CNN), are effective at picking up and exploiting small differences in data, especially for spatiotemporal coupled systems where the feature space is very large in higher order dimensions. As a result, improved performance can be achieved with access to higher dimensional data with finer temporal resolution and higher fidelity. Getting this data can be difficult for a tactical platform and using traditional computer modeling may not be sufficient due to its data approximations. However, scaled model data might be used to better capture the underlying dynamics and provide a critical element for the advancement of machine learning algorithms. Scale modeling cannot be a complete alternative and may be dismissed in the development and test of classification systems because of the expense when scaling to large class.


One other important consideration when using machine learning algorithms are generalization errors or systematic biases. Because these algorithms are sensitive to high dimensional features, they can often key on intangible artifacts like non-real sensor phenomenon or peculiarities present in the data collection. The traditional black box approach sometimes makes it difficult to detect or completely eliminate these types of errors; but all attempts must be made to do so. One way to do this is to ensure the algorithms are grounded in a priori knowledge of physical laws. As with human intelligence, machine intelligence must also be confined to the realm of reality.�


Recent mathematical tools have been developed that might be leveraged to resolve the apparent paradox of capturing the desired level of complexity in a machine learning algorithm and knowledge of the underlying physical mechanism it is exploiting. Examples of methods or techniques that may provide the desired results include the work by Raissi et al. [Refs. 2, 3], which has demonstrated the ability to translate noisy observations in space and time into non-linear partial differential equations. This was done by embedding a deep hidden physics layer in a Neural Network; it is able to learn the underlying dynamics during training [Ref. 2]. The resulting Neural Networks form the basis for new classes of algorithms with a priori built in knowledge of the underlying physical laws [Ref. 3]. This could allow better and more robust extrapolation to other objects within the same spatiotemporal framework using limited observations and/or augmented with computer and scaled model data. Another example of a technique used for complex dynamics is Dynamic Mode Decomposition [Ref. 5], which have shown the capability to extract governing equations of a dynamic system from sensor and image data collected on that system.


Combining new mathematical tools, hidden physics layers, scaled and computer models, and sparse observational data, it should be possible to build better and more robust intelligent machine learning algorithms. These new systems could process higher-dimensional input data at the same speeds or faster to achieve reduced missed identification or classification and increased correct identification and classification performance all the while providing higher confidence in those decisions. Existing data fusion metrics from Single Integrated Air Picture (SIAP) [Ref. 6] or the popular Stone Soup metrics package can be used to assess accuracy in identification and classification against existing systems as a baseline.


Work produced in Phase II may become classified. Note: The prospective contractor(s) must be U.S. owned and operated with no foreign influence as defined by DoD 5220.22-M, National Industrial Security Program Operating Manual, unless acceptable mitigating procedures can and have been implemented and approved by the Defense Counterintelligence and Security Agency (DCSA). The selected contractor and/or subcontractor must be able to acquire and maintain a secret level facility and Personnel Security Clearances. This will allow contractor personnel to perform on advanced phases of this project as set forth by DCSA and NAVAIR in order to gain access to classified information pertaining to the national defense of the United States and its allies; this will be an inherent requirement. The selected company will be required to safeguard classified material IAW DoD 5220.22-M during the advanced phases of this contract.


PHASE I: Design and develop a plan for implementing physics-based machine learning using sparsely sampled and noisy scaled laboratory data. Demonstrate feasibility of a sufficiently robust system to handle, and complex enough to leverage, spatial and temporal coupling and dynamic motion. The Phase I effort will include prototype plans to be developed under Phase II.


PHASE II: Based upon the plan from Phase I, develop a machine-learning classification algorithm for multiple targets with separate quarantined targets. The targets can be any class with spatial and temporal dynamics. Build a well-trained SVM, LSTM RNN, CNN classifier using physics-based hidden layers and scale model representations. Test and demonstrate the extent to which sparsely and/or noisy data from the quarantined target can be incorporated into the existing classifier. Test and demonstrate the extent to which the trained hidden physics layer can produce representative data that matches existing computer or scaled model data. Demonstrate the ability to generate data or a model that is robust against a well trained SVM or CNN classifier. The performance of the developed algorithm may be tested on an approved data set for validation.


Work in Phase II may become classified. Please see note in the Description section.


PHASE III DUAL USE APPLICATIONS: Extend the work to include real world data and accurate representative models. Transition the algorithm to appropriate military and commercial users. Heavy commercial investments in machine learning and artificial intelligence will likely continue for the near future. Better and more robust machine learning signal processing and classification has a myriad of commercial uses including financial market prediction, self-driving cars, medicine, and environmental research.



1. Abdulle, A., Weinan, E., Engquist, B. & Vanden-Eijnden, E. �The heterogeneous multiscale method.� Acta Numerica, 21, 2012, pp. 1-87. doi:10.1017/S0962492912000025


2. Raissi, Maziar. �Forward-Backward Stochastic Neural Networks: Deep Learning of High-dimensional Partial Differential Equations.�� Division of Applied Mathematics, Brown University, 2018.


3. Raissi, M. �Deep hidden physics models: Deep learning of nonlinear partial differential equations.� Division of Applied Mathematics, Brown University, 2018. ArXiv:1801.06637 [Cs, Math, Stat].�


4. �2018 National Defense Strategy.� United States Congress.


5. Manohar, K., Kaiser, E., Brunton, S. L. and Kutz, J. N. �Optimized Sampling for Multiscale Dynamics.� Multiscale Modeling & Simulation, 17:1, 2019, pp. 117-136.


6. Votruba, P., Nisley, R., Rothrock, R. and Zombro, B. �Single Integrated Air Picture (SIAP) Metrics Implementation.� Single Integrated Air Picture Systems Engineering Task Force, 2001.�


KEYWORDS: Scale Model, Machine Learning, Hidden Physics Layers, Non-Linear Differential Equation, Advanced Mathematics, Automatic Target Qualifier



The Navy Topic above is an "unofficial" copy from the overall DoD 20.2 SBIR BAA. Please see the official DoD DSIP Topic website at for any updates. The DoD issued its 20.2 SBIR BAA on May 6, 2020, which opens to receive proposals on June 3, 2020, and closes July 2, 2020 at 12:00 noon ET.

Direct Contact with Topic Authors. During the pre-release period (May 6 to June 2, 2020) proposing firms have an opportunity to directly contact the Technical Point of Contact (TPOC) to ask technical questions about the specific BAA topic.

Questions should be limited to specific information related to improving the understanding of a particular topic�s requirements. Proposing firms may not ask for advice or guidance on solution approach and you may not submit additional material to the topic author. If information provided during an exchange with the topic author is deemed necessary for proposal preparation, that information will be made available to all parties through SITIS (SBIR/STTR Interactive Topic Information System). After the pre-release period, questions must be asked through the SITIS on-line system as described below.

SITIS Q&A System. Once DoD begins accepting proposals on June 3, 2020 no further direct contact between proposers and topic authors is allowed unless the Topic Author is responding to a question submitted during the Pre-release period. However, proposers may submit written questions through SITIS at, login and follow instructions. In SITIS, the questioner and respondent remain anonymous but all questions and answers are posted for general viewing.

Topics Search Engine: Visit the DoD Topic Search Tool at to find topics by keyword across all DoD Components participating in this BAA.

Help: If you have general questions about DoD SBIR program, please contact the DoD SBIR Help Desk at 703-214-1333 or via email at [email protected]