Russian engineers have developed and tested a new drone. He easily detects a dangerous plant - hogweed
Sosnovsky's hogweed is a large herbaceous plant, a species of the genus hogweed of the umbrella family.
The sap of the plant, getting on the skin, under the influencesunlight causes severe burns. Moreover, their treatment requires close medical supervision for several weeks. Note that now the spread of Sosnovsky's hogweed has become a real ecological disaster - it has spread from the central part of Russia to Siberia, from Karelia to the Caucasus.
Sosnovsky's hogweed is one of the most famous and problematic invasive species in Russia.
One of the problems in dealing with him is hisexceptional vitality and full seed distribution. To find it, you have to manually go around the fields or use flying machines. Unfortunately, most satellites are unable to provide high enough resolution to detect single plants. At the same time, the accounting of plants using UAVs is not sufficiently automated and is often based on the use of aircraft that are expensive to operate.
Input image (left) and the result of the proposed fully convolutional neural network (right)
To fix the problem, the researchers tookdecision to use UAV. They are special in that they provide the latest information on the distribution of a plant with an exceptionally high resolution, even when the sky is covered with clouds.
As a hardware platform, they chosea DJI Matrice 200 quadcopter and an NVIDIA Jetson Nano single board computer with a relatively powerful video accelerator that allows you to launch directly on a neural network device.
Orthophotomap with the areas of hogweed growing marked on it (in bright green)
A convolutional neural network (CNN) is responsible for searching for a hogweed in frames from a drone camera, which receives a frame and carries out semantic segmentation, marking areas with a hogweed on it.
Recall that a convolutional neural network is a specialArtificial neural network architecture, proposed by Jan Lekun in 1988 and aimed at efficient pattern recognition, is part of deep learning technologies.
Developers have chosen three popular architecturesCNN to compare their performance for this task: U-Net, SegNet and RefineNet. The researchers themselves put together a dataset to train the algorithms. To do this, they shot many drone footage in the Moscow region, using two different drones and one action camera (attached to the drone). As a result, 263 images were obtained, in which the authors of the development marked the areas with hogweed. The dataset itself is available on GitHub.
Having trained neural networks, the authors tested them onsingle-board computer and found out that they operate at a frequency of tenths or hundredths of a frame per second. The best result was given by a network based on U-Net - 0.7 frames per second. The best classification was shown by a SegNet-based network with an area under the ROC-curve (a common metric for assessing the quality of a binary classification) equal to 0.969.
Nuclear reactions intensified in the reactor of the Chernobyl nuclear power plant
Scientists have shown how a black hole tears apart a star
Physicists have created an analogue of a black hole and confirmed Hawking's theory. Where it leads?