Study on Tobacco Plant Cross-Level Recognition in Complex Habitats in Karst Mountainous Areas Based on the U-Net Model

The extraction of crop information is one of the important research directions for precision agriculture remote sensing. Crop extraction is of great significance in crop refinement management, precision fertilization, growth monitoring and yield precision estimation. The karst mountainous areas in southern China are characterized by undulating terrain, broken cultivated land, scattered spatial distribution of tobacco planting plots, uneven growth of plants, and mixed planting of crops. As the flight height of UAVs increases, the area of tobacco planting plots decreases, and the texture features become increasingly blurred, which increases the difficulty of segmentation and affects the recognition accuracy. We conducted this study to explore whether the high-resolution sample datasets and the trained U-Net model are suitable for cross-level recognition. In this study, DJI Mavic 2 Pro was used to collect UAV RGB images with flight heights of 50 m, 60 m, 70 m and 90 m in complex habitats for extracting tobacco plants from the U-Net model. The results are as follows: (1) The precision of tobacco plant segmentation at different altitudes is 50 m > 60 m > 70 m > 90 m, and Kappa coefficient is 0.92, 0.89, 0.86 and 0.34; the pressure is 0.96, 0.94, 0.93 and 0.22; the recall is 0.91, 0.90, 0.86 and 0.24; and the IoU is 0.88, 0.85, 0.8 and 0.23, respectively; and the precision of complex background segmentation is: a small number of weeds > a large number of weeds, and the plot is flat > the plot is broken. (2) With increasing flight height, the precision of tobacco segmentation of the U-Net model gradually decreases. Compared with 50 m, the precision of the 60 m segmentation results is reduced by 0.03, 0.02, 0.01 and 0.03, and that of 70 m is reduced by 0.06, 0.03, 0.05 and 0.08. The precision of the 90 m segmentation results is reduced by 0.58, 0.74, 0.67 and 0.65. The flight heights of 50 m, 60 m and 70 m have good experimental results, but the precision of 90 m segmentation is poor. The precision is mainly affected by the two factors of floor height and light. This study verified the feasibility and reliability of the high-precision extraction of tobacco plants at different altitudes by U-Net in complex habitats and has a certain reference value for research on the methodology and technical system of the deep learning recognition of crops in complex habitats in karst mountains.