CPVF: vectorization of agricultural cultivation field parcels via a boundary–parcel multi-task learning network in ultra-high-resolution remote sensing images

Abstract

Accurate recognition and vectorization of agricultural cultivation field parcels (CFP) are crucial for agricultural monitoring. However, the diverse sizes and shapes of parcels, inherent blurriness of boundaries, and adhesion of densely distributed parcels pose considerable challenges in extracting complete and separable parcels from high-resolution imagery. To address these issues, we propose an end-to-end cultivated parcel vectorization framework (CPVF) based on a boundary-parcel multi-task learning model. The CPVF comprises two components: the model introduced in this paper for CFP extraction, termed the drone-based cultivation parcel extraction multitask learning model (DCP-MTL), and the universal vectorization module (UVM) for post-processing. The model combines region, boundary, and distance tasks with a discrete cosine transform module for frequency domain feature extraction and an ensemble decoding block. The ensemble decoding block with deep-supervision, enhancing parcel region separability and boundary connectivity in complex and densely packed parcel scenarios. The UVM incorporates region–boundary interaction and topological relation-based hanging line extension to repair broken boundaries. Experiments on a newly developed the first large-scale ultra-high-resolution (UHR) dataset show that our method achieves a region IoU of 92.88 %, boundary IoU of 60.94 %, and over-segmentation and under-segmentation rates of 17.5 % and 20.7 %, respectively. The proposed method outperforms BsiNet by improving region and boundary IoU by 9.08 % and 20.6 %, respectively, and reducing over- and under-segmentation by 5.7 % and 7 %. We assessed the model’s transferability across ten regions and various farmland landscapes, demonstrating stable generalization. Ablation studies and comparisons confirmed that CPVF provides precise and effective CFP vectorization in diverse and complex farmland scenarios.