HH
H. Heinczinger
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
3 records found
1
Meteorite falls can occur at any given place on Earth, and instrument observations of their fireballs can help constrain search areas to recover freshly fallen fragments. Rapid recovery is essential in European field settings, where exposure to soil, moisture, and vegetation can quickly affect meteorite preservation. Drone-assisted aerial surveys of strewn fields combined with machine learning can offer a novel approach to assist in search and recovery efforts. Other studies have explored drone-based meteorite detection as well [1-4], yet existing approaches are often tied to specific terrains and are difficult to adapt to the diverse land-cover types that we encounter across Europe.
Here we present a flexible training-data generation strategy and detection pipeline that leverages transfer learning with a pre-trained Convolutional Neural Network (CNN) architecture to identify fusion-crusted meteorites. Museum specimens were photographed from multiple orientations on an automated rotating stage and composited onto representative ground backgrounds to synthesize a varied training set. After fine-tuning the network on this dataset, we validated it on drone imagery containing hidden meteorites in grassy environments at multiple flight levels and in a demarcated search area that was surveyed by drone. The system consistently detected meteorites across these controlled field tests, while also highlighting the importance of tuning image scale, flight altitude, and detection-window size.
As our approach to training data generation is adaptive, we can tune the pipeline to strewn field-specific land cover types, providing a practical route toward deployment in future search campaigns. With the pipeline tested in a controlled field setting, we now await the next meteorite-dropping fireball event in the Netherlands or surrounding countries to conduct real-world field trials to validate this novel airborne detection strategy. ...
Here we present a flexible training-data generation strategy and detection pipeline that leverages transfer learning with a pre-trained Convolutional Neural Network (CNN) architecture to identify fusion-crusted meteorites. Museum specimens were photographed from multiple orientations on an automated rotating stage and composited onto representative ground backgrounds to synthesize a varied training set. After fine-tuning the network on this dataset, we validated it on drone imagery containing hidden meteorites in grassy environments at multiple flight levels and in a demarcated search area that was surveyed by drone. The system consistently detected meteorites across these controlled field tests, while also highlighting the importance of tuning image scale, flight altitude, and detection-window size.
As our approach to training data generation is adaptive, we can tune the pipeline to strewn field-specific land cover types, providing a practical route toward deployment in future search campaigns. With the pipeline tested in a controlled field setting, we now await the next meteorite-dropping fireball event in the Netherlands or surrounding countries to conduct real-world field trials to validate this novel airborne detection strategy. ...
Meteorite falls can occur at any given place on Earth, and instrument observations of their fireballs can help constrain search areas to recover freshly fallen fragments. Rapid recovery is essential in European field settings, where exposure to soil, moisture, and vegetation can quickly affect meteorite preservation. Drone-assisted aerial surveys of strewn fields combined with machine learning can offer a novel approach to assist in search and recovery efforts. Other studies have explored drone-based meteorite detection as well [1-4], yet existing approaches are often tied to specific terrains and are difficult to adapt to the diverse land-cover types that we encounter across Europe.
Here we present a flexible training-data generation strategy and detection pipeline that leverages transfer learning with a pre-trained Convolutional Neural Network (CNN) architecture to identify fusion-crusted meteorites. Museum specimens were photographed from multiple orientations on an automated rotating stage and composited onto representative ground backgrounds to synthesize a varied training set. After fine-tuning the network on this dataset, we validated it on drone imagery containing hidden meteorites in grassy environments at multiple flight levels and in a demarcated search area that was surveyed by drone. The system consistently detected meteorites across these controlled field tests, while also highlighting the importance of tuning image scale, flight altitude, and detection-window size.
As our approach to training data generation is adaptive, we can tune the pipeline to strewn field-specific land cover types, providing a practical route toward deployment in future search campaigns. With the pipeline tested in a controlled field setting, we now await the next meteorite-dropping fireball event in the Netherlands or surrounding countries to conduct real-world field trials to validate this novel airborne detection strategy.
Here we present a flexible training-data generation strategy and detection pipeline that leverages transfer learning with a pre-trained Convolutional Neural Network (CNN) architecture to identify fusion-crusted meteorites. Museum specimens were photographed from multiple orientations on an automated rotating stage and composited onto representative ground backgrounds to synthesize a varied training set. After fine-tuning the network on this dataset, we validated it on drone imagery containing hidden meteorites in grassy environments at multiple flight levels and in a demarcated search area that was surveyed by drone. The system consistently detected meteorites across these controlled field tests, while also highlighting the importance of tuning image scale, flight altitude, and detection-window size.
As our approach to training data generation is adaptive, we can tune the pipeline to strewn field-specific land cover types, providing a practical route toward deployment in future search campaigns. With the pipeline tested in a controlled field setting, we now await the next meteorite-dropping fireball event in the Netherlands or surrounding countries to conduct real-world field trials to validate this novel airborne detection strategy.
Meteorites are valuable resources in planetary geoscience, since they provide invaluable insights into the composition and origin of planetary bodies, enhancing our understanding of planet formation and the solar system. Despite occurrences of meteorite-dropping fireballs observed every 2 years, traditional field search campaigns in the Netherlands have faced challenges in retrieving fragments. In this presentation we will outline an innovative approach for meteorite recovery through the integration of drones and machine learning algorithms. To enhance search efforts, drones can be used to gain access to restricted fields and increase efficiency. The proposed strategy employs Convolutional Neural Networks (CNNs), a type of machine learning algorithm, to analyze drone imagery and identify potential meteorites. In order to train the artificial neural network, we created an extensive library of image data of fusion-crusted meteorites from the Naturalis meteorite collection placed on a backdrop of various field and soil conditions. The algorithm was tested during a field experiment at Unmanned Valley, Valkenburg, South Holland, using real meteorites and ‘meteorwrongs’ (objects resembling meteorites). During the drone survey we captured images of the target area and then implemented our detection strategy based on machine learning to the new image data set. In this presentation we will discuss the results and future steps.
...
Meteorites are valuable resources in planetary geoscience, since they provide invaluable insights into the composition and origin of planetary bodies, enhancing our understanding of planet formation and the solar system. Despite occurrences of meteorite-dropping fireballs observed every 2 years, traditional field search campaigns in the Netherlands have faced challenges in retrieving fragments. In this presentation we will outline an innovative approach for meteorite recovery through the integration of drones and machine learning algorithms. To enhance search efforts, drones can be used to gain access to restricted fields and increase efficiency. The proposed strategy employs Convolutional Neural Networks (CNNs), a type of machine learning algorithm, to analyze drone imagery and identify potential meteorites. In order to train the artificial neural network, we created an extensive library of image data of fusion-crusted meteorites from the Naturalis meteorite collection placed on a backdrop of various field and soil conditions. The algorithm was tested during a field experiment at Unmanned Valley, Valkenburg, South Holland, using real meteorites and ‘meteorwrongs’ (objects resembling meteorites). During the drone survey we captured images of the target area and then implemented our detection strategy based on machine learning to the new image data set. In this presentation we will discuss the results and future steps.
Meteorites are valuable resources in planetary geoscience, since they provide invaluable insights into the composition and origin of planetary bodies, enhancing our understanding of planet formation and the solar system. Despite occurrences of meteorite-dropping fireballs observed every 2 years, traditional field search campaigns in the Netherlands have faced challenges in retrieving fragments. In this presentation we will outline an innovative approach for meteorite recovery through the integration of drones and machine learning algorithms. To enhance search efforts, drones can be used to gain access to restricted fields and increase efficiency. The proposed strategy employs Convolutional Neural Networks (CNNs), a type of machine learning algorithm, to analyze drone imagery and identify potential meteorites. In order to train the artificial neural network, we created an extensive library of image data of fusion-crusted meteorites from the Naturalis meteorite collection placed on a backdrop of various field and soil conditions. The algorithm was tested during a field experiment at Unmanned Valley, Valkenburg, South Holland, using real meteorites and ‘meteorwrongs’ (objects resembling meteorites). During the drone survey we captured images of the target area and then implemented our detection strategy based on machine learning to the new image data set. In this presentation we will discuss the results and future steps.
...
Meteorites are valuable resources in planetary geoscience, since they provide invaluable insights into the composition and origin of planetary bodies, enhancing our understanding of planet formation and the solar system. Despite occurrences of meteorite-dropping fireballs observed every 2 years, traditional field search campaigns in the Netherlands have faced challenges in retrieving fragments. In this presentation we will outline an innovative approach for meteorite recovery through the integration of drones and machine learning algorithms. To enhance search efforts, drones can be used to gain access to restricted fields and increase efficiency. The proposed strategy employs Convolutional Neural Networks (CNNs), a type of machine learning algorithm, to analyze drone imagery and identify potential meteorites. In order to train the artificial neural network, we created an extensive library of image data of fusion-crusted meteorites from the Naturalis meteorite collection placed on a backdrop of various field and soil conditions. The algorithm was tested during a field experiment at Unmanned Valley, Valkenburg, South Holland, using real meteorites and ‘meteorwrongs’ (objects resembling meteorites). During the drone survey we captured images of the target area and then implemented our detection strategy based on machine learning to the new image data set. In this presentation we will discuss the results and future steps.