Edge-Preserving Texture-Based Semantic Segmentation for Ultrahigh Resolution Images in Agricultural Scene

The rich textural information in images captured by unmanned aerial vehicles (UAVs) with ultrahigh spatial resolution has emerged as a crucial characteristic for identifying different land covers. Therefore, it is essential to develop texture-based semantic segmentation methods for these images, especially those of agricultural scenes. In this work, we propose a texture-based semantic segmentation method for ultrahigh-resolution images in agricultural scenes, which incorporates an encoding layer to capture textural information. Furthermore, we implement two modifications to refine the results. Firstly, to solve the problem of over-smoothed results with blurred edges, we employ a layerwise loss and reduce the number of layers in the backbone to enhance the role of the shallow layer in preserving edges. Secondly, a codeword-based nonlocal noise suppression module is designed to tackle salt-and-pepper noisy patches in large objects. To validate the proposed method, two UAV datasets obtained from an agriculture-intensive area are utilized for evaluation. The results demonstrate that our proposed method effectively learns textural features and accurately identifies multiple agricultural and garden plot types. In comparison to recent works, our approach outperforms them in terms of intersection-over-union (IoU) and edge-based IoU. Moreover, our method exhibits strong advantages in noise suppression and edge preservation.