Study On Multi-scale Identification Of Tobacco In Complex Habitat In Karst Mountainous Areas Using Low Altitude Remote Sensing

Accurately grasping the quantity and spatial distribution information of tobacco plants can provide a basis for fertilizer application,irrigation,and pest control in the later stages of tobacco plants.Timely and accurate acquisition of tobacco plant information is of great significance for agricultural precision management and yield estimation.Remote sensing technology is one of the effective means to map the spatial distribution of tobacco.Small and low-cost unmanned aerial vehicles provide a good remote sensing platform for obtaining data in Karst mountainous areas with broken ground and complex habitats,creating new possibilities for accurate crop mapping.The cultivated land in the Karst mountainous areas is fragmented,and tobacco cultivation is characterized by spatial dispersion,uneven growth of plants,and mixed cropping of crops.At the same time,as the flying height of drones increases,the projected area of tobacco targets becomes smaller,the shape changes,and the texture features become increasingly blurred,increasing the difficulty of segmentation and affecting recognition accuracy.In order to explore whether the U-Net model for high-resolution tobacco sample dataset and training is suitable for multi-scale tobacco recognition in complex habitats and for image recognition at what flight heights,the study took the tobacco planting area of Beipanjiang Town,Zhenfeng County,Guizhou Province as the research area,and used Xinjiang Mavic 2 Pro to collect UAV images with flight heights of 50 m,60m,70 m,and 90 m under complex habitats in Karst mountainous areas,to construct different plot fragmentation Meteorological changes and tobacco datasets from different planting environments are used to train U-Net models to explore the factors affecting multi-scale remote sensing identification of tobacco plants in complex habitats.The results are as follows:(1)The study considers the impact of factors such as plot,planting environment,meteorological changes,and tobacco plant characteristics on tobacco plant recognition.Through analyzing the characteristics of each factor,tobacco recognition scenarios such as different plot fragmentation,planting habitat complexity,meteorological changes(sunny,cloudy),and tobacco plant characteristics(growth,color,etc.)are set up.A total of 6612 tobacco plants in various scenarios are manually labeled,and the images and samples are randomly segmented,Build into an initial dataset;By supplementing and expanding the original sample,a supplementary dataset and an expanded dataset are constructed.(2)Under complex planting environments,the segmentation accuracy of tobacco plants with different flight heights is 50 m>60 m>70m>90 m,and the Kappa coefficients are 0.92,0.89,0.86,and 0.34,respectively;The Precision is 0.96,0.94,0.93,and 0.22,respectively;Recall was 0.91,0.90,0.86,and 0.24,respectively;The Io U is 0.88,0.85,0.8,and 0.23,respectively.As flight altitude increases,the accuracy of U-Net model tobacco identification gradually decreases,and flight altitude has a significant impact on the accuracy of tobacco identification.Compared to 50 m,the accuracy of 60 m segmentation results,Kappa coefficient,Precision,Recall,and Io U,respectively,decreases by 0.03,0.02,0.01,and 0.03,while 70 m decreases by 0.06,0.03,0.05,and 0.08,respectively.The accuracy of 90 m segmentation results,respectively,decreases by 0.58,0.74,0.67,and 0.65.50 m,60m,and 70 m have good experimental results,However,the model of sample training is not suitable for extracting tobacco plants from 90 m images,which is mainly affected by two factors: layer height and illumination.At the same time,under different planting environments,the more complex the planting background,the worse the tobacco segmentation results.The complexity of the planting environment has a significant impact on tobacco segmentation.(3)Error and omission factor analysis of tobacco plant segmentation results.Shadows,light intensity,and blurring of tobacco plants cause tobacco plants to be omitted and have a significant impact;However,soil,rock,and mulch film only affect the extraction of tobacco plants under the background of no weeds and a small amount of weeds,with a small impact.The U-Net model mistakenly extracts tobacco plants,including weeds,corn plants,shrubs,barley plants,rocks,electric wires,white mulch,and soil.The most commonly mistakenly extracted tobacco plants are corn plants,weeds,shrubs,and barley plants,with similar colors and textures.Similar colors and textures have a significant impact on tobacco plant recognition.The research is oriented towards the needs of crop information interpretation from unmanned aerial vehicle remote sensing images,and proposes a process for constructing small sample datasets in karst mountainous areas,providing reference basis for the construction of complex mountain sensing image sample databases.The feasibility and reliability of the constructed tobacco sample dataset training U-Net model for high-precision extraction of tobacco plants from multiscale unmanned aerial vehicle images in complex habitats were verified,which has certain reference value for the research on the methodology and technical system of deep learning crop recognition in karst mountainous areas with complex habitats.