China's Crop Parcel Training Dataset (CCPTD)

Semantic segmentation of remote sensing imagery is a key method for land cover classification and mapping. In a typical remote sensing segmentation and mapping project, having more training samples generally helps the model better extract target features and achieve higher mapping accuracy. However, annotating remote sensing data is a time-consuming and labor-intensive task. The goal of strategic sampling is to minimize the number of training samples required while maintaining model performance, thereby reducing data annotation costs. This dataset supports a case study on strategic sampling methods for remote sensing semantic segmentation, specifically focusing on cropland parcel delineation. It includes GF-2 satellite imagery from five distinct agricultural regions across China. For each image, 20% of the area was selected, resulting in a dataset of over 12,000 annotated cropland parcel samples. Each sample includes three types of labels: parcel extent, parcel boundaries, and distance-to-boundary maps, enabling multi-task learning.