UAVs Remote Sensing Citrus Recognition Based On Deep Learning

As one of the world’s most important origins of citrus,China’s citrus planting area and citrus production both ranked first in the world by the end of 2019,which was an important economic source for southern fruit farmers,and it was also an important way for many states,prefectures and cities to develop characteristic agriculture,promote farmers to create wealth and increase their incomes,and achieve industrial prosperity and poverty alleviation.Since the 21st century,UAV remote sensing technology and target detection technology have developed rapidly in China,new forms of agricultural production which supported by UAV technology have become popular,and fruit recognition based on deep learning has also become a research hotspot.This experiment uses DJI Mavic Pro series UAV provided by the institute,taking Yichang mandarin orange ecological park in Hubei Province as the research area,and take the citrus fruits in this park as the research object,hope to achieve higher accuracy and efficiency to identify citrus fruits in natural scenes by using YOLOv3 algorithm,and conduct correlation analysis to reduce the prediction error and achieve the purpose of predicting yield.For this purpose,the following research is made:(1)Acquire the multi-angle and multi-type citrus images of UAV through manual operation by UAV,mainly includes high angle shot,side shot,yellow fruit,green fruit,overexposed fruit,dense and sparse fruit,etc.Using Label Image software for fruit target labeling,and make a self-made citrus data set with 2094 pictures.(2)Build YOLOv3 orange recognition model,pay attention to modify the category and set the model weight operation.During the training,adjust the exposure and saturation of the image at random,multi-scale training and batch training shall be carried out according to the requirements of YOLOv3 algorithm.Combining the outputs of the three levels as the final output,in this way,citruses with large and small targets can be extracted at the same time to ensure that the training model will applicable to input images of different scales.(3)Bring the test pictures into the trained orange recognition model for detection,and the obtained detection effect picture is presented.Select a reasonable model evaluation method and index to conduct accuracy analysis of the model,the main indicators are PR curve,IOU intersection ratio and mAP value.At the same time,the accuracy,efficiency and shortcomings of the YOLOv3 citrus recognition model are analyzed according to the index results and the effect diagram.(4)According to the problems presented in the test results,an improved non-maximum suppression algorithm,namely NMS algorithm is proposed,and the validity of the algorithm is verified by comparison experiments to solve the problem of model accuracy decline caused by invalid detection frames.At the same time,Making regression analysis on the fruit number identified by the algorithm,the actual number of citrus fruits corresponding to the image(the real number is obtained by visual inspection,including the partially complete single-sided fruit tree images)and the predicted fruit number of some complete single-sided fruit trees,the actual number of citrus fruits in a whole tree,and the corrected prediction fruit value function and yield estimation function are obtained.Through experiments,verification and analysis,the conclusions of this study are as follows:(1)The YOLOv3 citrus recognition model can still maintain a high accuracy when the recall rate is high,the mAP on the test set is up to 0.75429.The detection level of citrus images from different angles and scenes is almost at the level of human eyes,and the overall identification accuracy is maintained above 90%,which can be seen that this algorithm in this paper can effectively detect citrus fruits in complex environments,and it has strong robustness.The test speed of single image is stable within 3 s,so the model has high recognition efficiency.However,due to the advantages of YOLO series model for small target detection are not obvious,there are still some problems of missed and wrong detection.(2)By analyzing the test results,it is found that the NMS algorithm of this model has the phenomenon of detection frame redundancy.Therefore,experimental comparison between the improved NMS algorithm and the original algorithm for many times are conducted,it is found that the improved algorithm can effectively eliminate some redundant detection frames without deleting the correct detection frames,which is helpful to improve the efficiency and accuracy of model identification,therefore,the proposed improved NMS algorithm has certain feasibility.(3)Through regression analysis of the predicted and true values of citrus fruits,and compare the R2 obtained by the regression equation fitting,although there are some errors and contingencies due to the small amount of data,after using the equation,the fitting effect significantly increased,therefore,it can be seen that the function equation can realize the correction effect on the predicted fruit value,and at the same time,the total output of citrus in the park can be predicted by the equation.According to the research,under the rapid development of UAV remote sensing technology and deep learning target detection algorithm,the image of citrus fruit in the natural scene acquired by the UAV obtained a better recognition effect with the YOLOv3 algorithm,the improved non-maximum suppression algorithm proposed in this study has certain feasibility to reduce invalid detection frames and improve the recognition accuracy,at the same time,the corrected predictive value equation and yield estimation equation obtained by regression analysis can be used to estimate the total amount of citrus fruits in the research area,and apply the results to real-world citrus production,it plays a certain role in improving the cultivation management level and economic income of fruit farmers.