Image Processing And Classification Of Oilseed Rape And Weeds Based On Simulation Platform Of Unmanned Aerial Vehicle

The production of oilseed rape is important to China, however the existence of weeds in rape field is so severe that need to be solved. In this research, one kind of oilseed rape called ZS758 and four species of common weeds in the field of rape were grown to be photographed by low altitude multispectral remote sensing, low altitude machine vision and hyperspectral instrument. And the acquired photos were used to classify the oilseed rapes and weeds. The content of this study is summarized in brief:(1) The D90 SLR camera was carried in the simulative UAV(Unmanned Aerial Vehicle) platform to take RGB images from the oilseed rapes and weeds. In order to segmenting out the objects, the image mosaic algorithm and automatic identification were used. After calculating various vegetation indexes of each plant, to find out which system structure of deep Convolutional Neural Network was the optimal, several system structures were tried to classify oilseed rape and weeds. It came out that the performance of the CNN-1 which the amount of the convolution filters was 32/32/32 was the best, and the accuracy-test was 90.2%. The investigation about the influence of shooting height and the speed of the device to the classification accuracy was also be taken.(2)The ADC multispectral camera was carried in the simulative UAV(Unmanned Aerial Vehicle) platform to take multispectral images from the oilseed rapes and weeds. In order to segmenting out the objects, the image mosaic algorithm and automatic identification were used. After calculating various vegetation indexes of each plant, the accuracy-test outcomes of NIR/R and ratio characteristic components (including NIR/R,NIR/G,R/G) were both optimal. The outcome based on ELM model was the optimal solution using selected vegetation indexes which average prediction accuracy was 95.71%. The research about the classification based on textural features of each oilseed rape and weed was conducted, and the outcome based on BPNN model was the optimal solution which average prediction accuracy was 88.57%. The effect of classifying oilseed rapes and weeds based on vegetation indexes was better than textural features. The investigation about the influence of testing time, shooting height and the speed of the device to the classification accuracy was also be taken.(3)The classification model of oilseed rapes and weeds based on full-wave bands and characteristic bands of hyperspectral images were built respectively. And after research and comparison, de-trending, loading-weights and ELM were found to be the optimal pre-processing method, the optimal characteristics of band selection method and the optimal model separately. The prediction accuracies were both 100% based on full-wave bands and characteristic bands of hyperspectral images. The variation trend of classification accuracy appeared to be the trough of wave with timeline extending.