Problem Statement

BACKGROUND

Air warfare is the business of maintaining a higher situational awareness than the enemy. See: https://csbaonline.org/uploads/documents/Air-to-Air-Report-.pdf

This has been a fact since the earliest days of war in the air.

"The enemy must be surprised and attached at a disadvantage, if possible with superior numbers so the initiative was with the patrol... the combat must continue until the enemy has admitted his inferiority, by being shot down or running away."

Oswald Boelcke in the "Dicta Boelcke", a World War 1 era flight combat manual.

German aces of World War 2, such as Erich Hartmann (352 kills) and Gerd Barkhorn (302 kills) stressed what they referred to as "ambush tactics" in Europe. In the Pacific, American aces such as Richard Bong (40 kills) and Tommy McGuire (38 kills) developed virtually identical tactics.

A detailed analysis of 112 air combat engagements during the Vietnam War conducted by the US Air Force concluded that 80% of aircrew shot down were unaware of the impending attack.

"Despite vast changes in aircraft, sensor, communication, and weapon capabilities over the past century, the fundamental goal of air combat has remained constant: leverage superior [Situational Awareness] SA, to sneak into firing position, destroy the opposing aircraft, and depart before other enemy aircraft can react."

Trends in Air-to-air combat - Implications for Future Air Superiority, John Stillon, CSBA report.

Modern sensors primarily focus on radar detection of hostile aircraft, with a dearth of ground-based optical sensors.

PROBLEM:

There are several problems intended to be solved by this project:

1. In order to avoid being surprised by radar guided munitions, modern military aircraft usually come equipped with a radar warning receiver. This set of antennae and processing equipment passively "listen" for radio waves used by radar and provide the pilot with warnings on (i) the presence of a radar, (ii) when the radar is tracking the aircraft, and (iii) when the radar is guiding a missile onto aircraft.

Radars cannot be kept on indefinitely. Imagine a dark room. Turning the radar on is the equivalent of turning on a torch in the said dark room. Anti radiation missiles can home in on the radar emissions. Certain aircraft are also specially equipped to triangulate the position of an emitting radar to provide coordinates to other precision guided munitions (see: AN/ASQ-213 HARM targeting system)

Combining radar detection with other, passive sensors, permit greater SA than radar systems alone.

2. Missiles rockets do not burn the entire time of flight. They usually only have a few seconds of fuel before the missile relies on momentum to coast to the target (exceptions such as the MBDA Meteor exist, but that is an exception). This means that missiles can be kinematically defeated by simply turning in the other direction. The further away the target is when launch occurs, the more reaction time the pilot has.

A passive sensor would permit the radars to remain off until the target enters the no escape zone (where the missile will still catch the target if it turns and runs and there is accordingly a high Pk (kill probability)).

3. No sensor net is perfect, additional passive sensors that can be acquired cheaply (cameras and consumer GPUs) offers a capability boost with a small expenditure of resources.

OBJECTIVES:

1. Build a model which detects military aircraft from an image.

2. Build a second model to identify the type of aircraft detected by the first model.

3. Integrate both models into a program which take in (i) coordinates of camera, (ii) elevation of camera, (iii) altitude of camera, and (iv) information from the models, and return (i) type of aircraft, (ii) estimated coordinates of aircraft, and (iii) estimated altitude of aircraft.

SCOPE:

1. Use the data to feed various machine learning models to achieve the objectives.

2. Select the best detection model based on precision/recall, and the best classification model based on accuracy.

3. Develop program to satisfy the requirements of objective 3.

DATA:

1. Military aircraft detection dataset. https://www.kaggle.com/datasets/a2015003713/militaryaircraftdetectiondataset

METHODS AND TOOLS:

1. Neural networks including (i) ConvNext, (ii) EfficientNet, (iii) YOLOv8, (iv) RT-DETR.

SUCCESS METRICS:

1. Accuracy.

2. mAP-50.

Hardware and related requirements

CUDA compatible GPU: https://developer.nvidia.com/cuda-gpus

CUDA >= 11.8

Detector Notebook Requirements

cv2 4.8.1

matplotlib 3.7.2

numpy 1.25.2

pandas 2.1.0

sklearn 1.3.0

tqdm 4.66.1

ultralytics 8.0.207

yaml 6.0.1

Classifier Notebook Requirements:

keras 2.14.0

keras-core 0.1.7

#matplotlib 3.7.2

numpy 1.25.2

Function testing and development Notebook Requirements:

cv2 4.8.1

geographiclib 2.0

keras_core 0.1.7

matplotlib 3.7.2

numpy 1.25.2

pandas 2.1.0

ultralytics 8.0.207

py files requirements:

geographiclib==2.0

keras-core==0.1.7

numpy==1.26.2

opencv-python==4.8.1.78

pandas==2.1.3

pytz==2023.3.post1

torch==2.1.1

torchvision==0.16.1

ultralytics==8.0.218

Model Performance

Detector

Model	Epochs	Precision	Recall	mAP50	Remarks
RT-DETR	100	0.953	0.866	0.902	This performance is acceptable
RT-DETR	100	0.936	0.875	0.905	With additional augmentation
RT-DETR	100	0.953	0.849	0.889	With less augmentation
RT-DETR	100	0.96	0.868	0.902	Using cosine LR scheduler
YOLOv8	100	0.89	0.778	0.883	Somewhat worse performance

Classifier

Model	Optimiser	Train Acc	Train Loss	Val Acc	Val Loss	Fit time
ConvNext Base	AdamW	0.9811	0.0647	0.7831	1.1933	58m 29s
ConvNext Base	RMSProp	0.9760	0.0702	0.7766	1.3348	51m 11s
ConvNext Base	Adam	0.9761	0.0743	0.7724	1.1510	50m 41s
ConvNext Base	Adamax	0.9981	0.0330	0.8031	0.8546	1hr 12m 39s
ConvNext Base	SGD	0.9620	0.2507	0.7802	0.8217	12hr 17m 2s
ConvNext Base	Adadelta	0.9978	0.0091	0.9100	0.5438	76m 13s
ConvNext Base	Adagrad	0.0362	15.5349	0.0384	15.5020	52m 52s
ConvNext Base	Nadam	0.5144	1.6397	0.1886	4.7987	1hr 59m 7s
ConvNext Base	Ftrl	0.0619	3.6134	0.0553	3.6123	2hr 47m 16s

Program

The above has been consolidated into a single program with a GUI > run from interface.py.

Test images can be found in Assets.

A video of the program working can be found here: https://www.youtube.com/watch?v=MkIymKNswYQ

Conclusions and recommendations

1. The optical distributed IADs Network component system is theoretically feasible and successfully detected and located target aircraft.
2. It can be used to supplement traditional detection mechanisms to provide greater air defence coverage.