Research On Citrus Information Extraction From UAV Multispectral Image Based On Deep Learning

Citrus is an important economic agricultural crop extensively cultivated in the southern regions of China.It holds significant economic significance in terms of increasing farmers’income and contributing to the development of new rural areas.The rapid and accurate acquisition of cultivation information is a prerequisite for precise management of citrus orchards.Particularly,identifying and mapping the planting areas within the orchards can provide valuable distribution information about citrus trees.Accurately obtaining citrus tree information can serve as a reference for subsequent management and yield estimation.In the past,obtaining citrus tree information mainly relied on labor-intensive and time-consuming manual field measurements,which were inefficient.However,in recent years,with the development of deep learning and the application of unmanned aerial vehicle(UAV)remote sensing technology in the agricultural field,platforms and solutions have been provided for the rapid acquisition of citrus information.Liuzhou is one of the important citrus production areas in China,known for its abundant citrus resources.This study focuses on a citrus orchard in Luzhai County,Liuzhou City,and explores the extraction and application of citrus tree information using a citrus tree dataset generated from UAV images.The studied parameters include the number of citrus trees,individual tree canopy width,and the relationship between spectral information extraction and citrus yield.To improve the accuracy of citrus tree identification and counting in real-world scenarios,an improved version of YOLOv4(You Only Look Once v4),combined with the MobileNetV3 network,is proposed.This approach reduces the model’s size and improves detection speed.Additionally,the Convolutional Block Attention Module(CBAM)is employed to enhance the network’s feature extraction capability and is combined with the Adaptively Spatial Feature Fusion(ASFF)to enhance the network’s multi-scale feature fusion capability.Furthermore,a learning strategy based on cosine annealing is used during model training to accelerate the training speed and improve detection accuracy.A novel deep learning-based method,U~2-Net,is adopted for automatic extraction of citrus tree crowns.The proposed method is compared with three mainstream deep learning models(PSPNet,U-Net,and DeepLabV3+)in three typical experimental regions to extract citrus tree crown information,and the extraction results are compared and analyzed.Finally,an improved combination of YOLOv4,U~2-Net,and Arc GIS is used to extract citrus parameter information,which is then combined with four machine learning models(CNN,RBFNN,RFR,and SVR)to establish a yield prediction model.The main conclusions are as follows:(1)The improved YOLOv4 model effectively overcomes noise in the orchard environment,achieving fast and accurate identification and counting of citrus trees while reducing the model’s size.The crown detection mean average precision(mAP)reaches 95.86%,the frames per second(FPS)reaches 77.24,and the F1-score reaches 0.92.The average overall accuracy(AOA)for counting reaches 95.51%.In summary,the YOLOv4-MCA model meets the practical requirements of citrus tree detection and counting in this study,providing optional technical support for the digitalization,precision,and intelligence development of smart orchards.(2)Among the three experimental regions,the U~2-Net model achieves the highest accuracy in extracting citrus tree crowns,with intersection over union(IoU),overall accuracy(OA),and F1-score reaching 91.93%,92.34%,and 93.92%respectively.Compared to the other three deep learning models,the U~2-Net model improves IoU,OA,and F1-score by 3.63-8.31%,1.17-5.25%,and 1.97-4.91%respectively.Additionally,there is high consistency between the U~2-Net model’s extracted crown area and the measured area,with R~2 values above 0.93 for all three experimental regions.Compared to the other three deep learning models,the U~2-Net model also exhibits lower error rates,with an RMSE of 1.35m~2and an MRE of 8.15%.The results indicate that the combination of UAV multispectral images and the U~2-Net model enables accurate extraction of citrus tree crowns,with well-preserved crown contours,providing fundamental data and technical support for monitoring citrus growth dynamics and yield prediction.(3)This study accurately extracts individual citrus tree crown area,spectral features,and individual citrus fruit.Among all features,crown area,RVI,and NDVI contribute the most to citrus yield prediction.Therefore,these three features are selected for training the machine learning models.When using the selected three features,the CNN yield prediction model achieves a training set R~2 of 0.846 and a testing set R~2 of 0.734,outperforming the other three yield prediction models.