A Unified Framework for Cropland Field Boundary Detection and Segmentation

Rodrigo Fill Rangel, Vítor Nascimento Lourenço, Lucas Volochen Oldoni, Ana Flavia Carrara Bonamigo, Wallas Santos, Bruno Silva Oliveira, Mateus Neves Barreto; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2024, pp. 636-644

Agricultural monitoring is essential to ensure food security while minimizing the environmental impacts generated by the activity. Crop fields are the basic units of management in farmland, and the delimitation of their boundaries is useful for farmers and field-level analysis. In this work, we address the cropland field boundaries segmentation challenge by proposing an end-to-end novel segmentation framework. Our framework comprises three main pipelines: data preparation, in which Sentinel-2 MSI sensor images are handled; segmentation, where we propose the use of three different methods to obtain a cropland field map, the Felzenswalb's segmentation algorithm, and the neural networks U-Net and R2AttU-Net; and post-processing, where we propose a novel temporal aggregation and filtering methodology to enhance crop field boundary delineation. Our results show that our end-to-end framework is able to outline cropland field boundaries from Sentinel-2 data. The U-Net segmentation achieved overall good results, although some small fields may not be correctly identified. On the other hand, the post-processing was able to mitigate most incorrectly segmented cropland field polygons, significantly improving results in most metrics, removing isolated pixels, and better delimiting fields.