Finding Anti-Personel Mines

 The deployment of Russian PFM-1 anti-personnel mines (also known as butterfly or petal mines) in Ukraine poses a unique problem, these mines are quite small approximately 5 inches by 2 inches, and are scattered over an area by aircraft. The size and method of deployment mean that these mines often end up undetected and on civilian property.  The best way to mitigate any potential damage caused by these mines is to locate all mines in a given area. A proven method for finding buried mines is using thermal imaging to locate the temperature differential caused by the buried mine. In this project, I investigate the efficacy of Deep Learning models and how it compares to thermal imagery for the detection and location of these small aerially scattered anti-personnel mines.

• Locate sporting clays within mock minefields using deep learning
• Determine the ideal deep learning model to find the sporting clays
• Create an effective model
• Create a road map for others to use deep learning to detect objects
• Compare the efficacy of deep learning to thermal imagery

Equipment and software used in this project are as follows

• Drones: Skydio 2+, DJI M300
• Sensors: Skydio 2+(RGB), DJI Zenmuse H20t(RGB/Thermal)
• Landmine Stand-In: 270 Sporting Clays
• Photogrammetry Software: Pix4D Mapper, Agisoft Metashape
• GIS Software: ArcGIS Pro, ArcGIS Online
• Deep Learning: ArcGIS Pro

To collect imagery of a high enough resolution to be able to identify the sporting clays using deep learning an ideal GSD (Ground Sampling Distance) of <= 1 CM was identified as the sporting clays have a diameter of 10.8 CMs and a clear and defined shape is necessary for object detection. To achieve this GSD, imagery was collected at approximately 60 ft above ground level using the Skydio 2+ with an overlap of 85% on both the front and the sides. The imagery collected was processed using Pix4D Mapper.

To create an accurate deep learning model I need a diverse data set with a multitude of ground cover types to remove any potential biases that a highly monolithic data set could impart. The dataset I created to train my models included as much diversity of ground cover and positioning as possible within the area I was using.

To train my deep learning models I used the integrated tool inside ArcGIS Pro called "label objects for deep learning" to label objects in my training data set as my target objects. The output of this tool is what is then used to train a model using the tool "train deep learning model". For a more in-depth walkthrough click the button below.

Thermal imagery was collected using the DJI M300 with the DJI Zenmuse H20t thermal camera. Due to the constraints of the study area data had to be collected at 100 ft above ground level, however with the H20t sub-centimeter RGB was still collected. The overlaps for this collection were both set at 90%. The imagery collected was processed using Agisoft Metashape.

To assess how effective thermal would be at detecting of similar sizes to the Russian PFM-1 I conducted a small scale test. I set out 24 sporting clays on both grass and concrete and then captured imagery of both the layout of the sporting clays and the broader area using the DJI Zenmuse H20t thermal sensor.

The goal of this test was to identify how effective and if a larger scale test comparing thermal data with the final deep learning model developed to compare and contrast their effectiveness and efficiency was necessary. After completing this test and reviewing the data produced by it became evident that both the heat differential between the clays and their surroundings and the resolution of the thermal sensor made static thermal data less than adequate as a primary tool for finding objects of similar size to the sporting clays. These results did not make it necessary to perform a larger test.

• While not perfect Deep Learning Models provide a very effective and highly scalable method for detecting very small objects like PFM-1 landmines and many of the false positives and negatives that occur can be eliminated with larger and more diverse sample sizes.
• Although Single Shot Detectors may be faster RCNN models yield cleaner and more accurate results and will fail positive whereas Single Shot Detectors will fail both positively and negatively. I believe this makes RCNN models the better choice for this application.
• While it is possible to identify the sporting clays from thermal imagery it is difficult to determine what is and isn't the target without using RGB imagery as well. This makes the use of thermal imagery nowhere near as scalable as Deep Learning models, however, thermal imagery could be highly useful in locating actual mines to develop a training dataset for a practical deep learning model