Location Prediction and Image Recognition of Asian Hornet Based on Deep Learning

Abstract

AbstractIn recent years, the frequent occurrence of alien species invasion not only reduces the diversity of local species, but also greatly affects the production and life of human beings. To help the government solve the problem of Asian bumblebee invasion in the problem, this paper establishes a migration learning model of Long Short Term Memory neural network, convolution neural network and neural network to solve the location prediction of Asian bumblebee propagation and the recognition of Asian bumblebee in bee pictures. To determine the next propagation location of bumblebee in Asia, this paper first filters and sorts the attached data to obtain the time, latitude and longitude information Positive ID by 14 Lab State fields arranged in chronological order. LSTM neural network suitable for high precision prediction of small data sets is selected to solve the problem of small and 2 dimensional data sets. According to the data set size, the sliding window size is four, the first three data of each window are used as the training set, and the last data of the window is the most tested set. A LSTM neural network with a scale of 120 ~ 120 ~ 2 is established to train and learn the data set after grouping. The correct probability of prediction results is 98.41, so we think that the model can solve the problem of location prediction of Asian hornet propagation over time. Using the training results of the last three data as the prediction results, the latitude and longitude of the next propagation position is obtained as [48.94857702, −122.581335]. The distance between the predicted position and the latest data position is 3.44 km, which is smaller than the maximum active radius of Asian bumblebee 8 km, so it is considered that the predicted data are reasonable. To identify Asian bumblebees in bee pictures, this paper filters the attached data and obtains pictures with Positive ID and Negative ID fields. Since there are only 14 Positive ID data sets, CNN networks with high image recognition accuracy are selected. Then separate the two types of pictures and number them separately, and the image of 128 * 128 into 128 * 128 * 3 pixel matrix, enter into the neurons of the CNN network for training and learning, 10 rounds of training, 120 sessions per round, each time 20 samples were randomly selected for training, of which 80% were randomly selected as training data, 20% for test data. Finally, the accuracy of the training set is 99.67%, the accuracy of the test set is 99.55. Therefore, we think that the model can solve the problem of Asian bumblebee recognition in bee pictures.KeywordsLong and short term neural networkConvolution neural networkTransfer learning