The Research on Precise Monitoring Methods for Grain Planting Areas Based on High-precision UAV Remote Sensing Images

Precisely monitoring the range of rice cultivation is an essential task for the government to dynamically supervise the red line of 180 million mu ($ 1 mu \approx 666.667 {m^2} $) of arable land. This study aims to address the issues of low efficiency, high cost, and insufficient accuracy in traditional rice cultivation range monitoring methods. Against the backdrop of the widespread application of UAV remote sensing and the maturity of deep learning technology, this paper constructs a high-precision UAV remote sensing image dataset for rice identification, which includes different growth stages of rice, different resolutions, and regions. It also utilizes deep learning semantic segmentation technology to study the models, remote sensing image resolutions, and model sample sizes suitable for precise monitoring of rice. The experimental results show that, on the basis of balancing cost, efficiency, and accuracy, the Deeplabv3+ and PSPNet models combined with remote sensing image data of 8 cm resolution are more suitable for monitoring and extraction of rice cultivation areas, and PSPNet has a stronger few-shot learning ability. In response to the strong model generalization ability under the dispersed rice cultivation areas and diversified features, this paper proposes a method of transfer learning with a small number of samples. This method has a more stable training process, and the IoU is 5% $\sim$ 10% higher than that of unsupervised transfer learning models and fully supervised models with a small number of samples.