Uav Image Recognition Of Mangrove Based On Convolutional Neural Network

Mangrove habitat is unique,species diversity,obvious ecological function,good social benefits,with high ecological protection value.In recent years,due to the overdevelopment of land,the mangrove area is also increasingly reduced,mangrove protection has become extremely urgent and important.It is of great significance to study an efficient method for mangrove monitoring and provide a theoretical basis for mangrove conservation.At present,most of the conventional research methods on the classification of mangrove species require manual feature extraction.In addition,some scholars adopted the method of the fully connected neural network to identify and monitor mangroves.However,the fully connected neural network needs to train a lot of parameters because the neurons in each layer of the network are connected in a full-connection way,and the learning time is too long,and the purpose of convergence may not be achieved.In recent years,deep learning has developed rapidly and made important progress in the field of image recognition.As a major deep learning method,convolutional neural network(CNN)has significantly improved the image feature extraction and training effect compared with the fully connected neural network due to the addition of convolutional layer in the network model.At present,when referring to the related literature of mangrove species monitoring,it has not been found that researchers use convolutional neural network method to identify mangrove species.Therefore,it is of great significance to study the identification of mangrove species by means of convolutional neural network.The main research process is as follows:1、picture collectionIn this paper,shenzhen bay mangrove reserve was selected as the research area,and dji spirit 4Pro uav was used as the mangrove image acquisition platform to collect high-quality mangrove images in the research area.The image acquisition scheme is as follows:the height of image acquisition is set as 100m,the flight route of the uav is set as s-type,and the flight speed is 3m/s.A picture is taken every 1 second,and the coincidence of the pictures is about 1/3.2、Data set making and network model designThe collected mangrove graphs were spliced into a complete mangrove monitoring map in the study area,and the spliced images were cut into images with a size of 28×28.3,200 images of bones soil,1,900 images of red sea olive,3,200 images of wood olive,3,200 images of autumn eggplant and 3,200 images of background were selected for marking.1,300 images were selected from red sea olive and flipped to expand the data,bringing the number of red sea olive markers to 3,200.The 16000 sample data marked in five categories were randomly scrambled to make a data set,which was divided into training set,verification set and test set in a ratio of 3:1:1 for model training,verification and test.The structure and characteristics of LeNet-5 network model and AlexNet network model were analyzed,and appropriate improvements were made to the network model LeNet-5(1)、LeNet-5(2)and AlexNet(1)for mangrove species identification.3、Model training and test result analysisSet model training parameters,where the learning rate is set to 0.001,0.0001 and 0.00001 to train each model.The main analysis indexes are the overall test accuracy of mangrove species,the test accuracy of each species and the test accuracy volatility of each species.The analysis index selected the best model for the prediction of mangrove species.4、Mangrove species area calculation and species distribution showAll mangrove sample images were input into the prediction model for species prediction,and the predicted number of various mangrove species was counted.The area of various mangrove species was calculated according to the imaging principle of the camera and relevant parameters.The predicted mangrove species were color converted,and the distribution information of various mangrove species was spliced.Based on the above studies,LeNet-5(1)with Leaky-ReLU activation function and a learning rate of 0.0001 had the best test effect on mangrove species.The overall test accuracy of species was the highest,reaching 0.91.The test accuracy of baigurang,red sea olive,magnolia and autumn eggplant reached 0.92,0.95,0.91 and 0.85 respectively.Their calculated areas are 1058m~2,300 m~2,72m~2 and 910m~2 respectively.Color conversion and splicing of predicted mangrove species images to display mangrove species distribution information.This research method has realized the high resolution identification of mangrove species and provided important theoretical basis for the information monitoring and species protection of mangrove ecosystem.