Abstract: Unmanned aerial vehicle (UAV) remote sensing is widely used in land cover classification, but at present, object recognition of UAV images in farmland environment is cumbersome, requiring manual image stitching to preprocess the image, and finally training the model for recognition. In order to improve the computing speed and efficiency of UAV images in farmland environment, this paper proposes a parallel Mosaic recognition algorithm for UAV images in farmland environment. Firstly, based on SURF algorithm, KNN algorithm and RANSAC algorithm, the panoramic image Mosaic of UAV based on image transformation is realized. Then the inverted binary tree parallel processing algorithm is proposed. The parallel processing algorithm is planned by the idea of divide and conquer, and the image Mosaic recognition task is automatically divided into multiple sub-processes according to the number of CPU cores of the edge device and the number of images. And they are distributed to different computing cores to execute, and to improve the computing efficiency in the field environment. Finally, by extracting the transformation matrix in image Mosaic, the recognition was realized at the same time in the process of image Mosaic, and the recognition results were also spliced by using the transformation matrix. The parallel splicing recognition algorithm greatly improves the efficiency of UAV image splicing recognition, which is of great significance to realize real-time monitoring of UAV images. The experimental data were taken in the Science and Education Park of Henan Agricultural University and the Pest Field Observation Station base of the Ministry of Agriculture and Rural Affairs in Korle, Yuanyang County, Henan Province. In order to increase the diversity of data, Phantom4RTK and Mavic3E were used to avoid the influence of different Uavs on the Mosaic effect. The collected images are RGB three-band images, and the shooting height is about 20-100 m. The resolution of the Sprite 4RTK camera is 5472×3648 pixel, and the resolution of the Imperial 3E camera is 5 280×3 956. The CPU model of the equipment selected for the experiment is i7-11700, the frequency is 2.50 GHz, there are 8 cores and 16 logical processors, and the memory is 64GB. The graphics card of the model training device is NVIDIA A100-PCIE, the neural network is U-Net, and the development framework is Pytorch. The training epoch is 300 rounds, and the backbone network is frozen for training in the first 50 rounds to avoid too much data noise. The batch size (batch_size) is set to 16 and the loss function is the cross entropy loss (CE). Model training significantly reduces the loss value after the start of training and converges at 150 rounds and becomes stable at 200 rounds. The final mIoU of the model training was 89.01%. We use the reverse binary tree parallel stitching algorithm and other stitching algorithms for comparative experiments. The experimental results show that under the same experimental environment and data set, using the inverted binary tree parallel stitching scheme, the stitching time is reduced by about 60%-90% on average compared with commercial software. We test the inverted binary tree parallel splicing recognition algorithm in a farmland environment, and deploy the initially trained model to the device. The experimental results show that compared with the serial Mosaic recognition, the parallel Mosaic recognition of inverted binary tree not only reduces the time consumption by 70%, but also improves the mIoU of image recognition by 10.17%. It shows that in the farmland environment, the multi-threaded inverted binary tree parallel scheme can make better use of the computing resources of edge devices in the farmland environment, greatly improve the speed of UAV image stitching and recognition, and provide technical support for rapid real-time monitoring of UAV.