Integration of harvester trajectory and satellite imagery for large-scale winter wheat mapping using deep positive and unlabeled learning

Abstract

Limited accurate ground truth labels are the primary constraint for data-driven modeling analysis of large-scale crop mapping. Existing labeling methods largely rely on field surveys, visual interpretation, and historical ground information. These labor-intensive approaches are often limited by spatiotemporal heterogeneity of crop distribution and encounter the challenge of gathering extensive crop labels. The massive operating trajectories of agricultural machinery contain precise location information of the crop fields, providing a new source for accurate crop labels at a large spatial scale. This study develops a large-scale crop mapping workflow through widespread harvester trajectory and 10 m Sentinel-2 imagery. The trajectory-based automatic labeling method is developed to generate 287,533 winter wheat labels by jointly using harvester coordinates and satellite images. These generated one-class ground labels are further used to develop positive and unlabeled learning based deep learning models for winter wheat mapping. The Positive and Unlabeled Learning-based Convolutional Neural Network (PUL-CNN) outperforms the other four one-class based classifiers with an F1 score of 94.4 % at 12 study sites. The estimated county-level winter wheat acreage agrees well with census data with R2 of 0.86 in the overall study region. The interpretation analysis based on the Shapley Additive Explanation method shows the heading and greening stages are the critical periods for wheat mapping, aligning well with the separability in Normalized Difference Vegetation Index (NDVI) curves. The results of winter wheat mapping demonstrate the integration of harvester trajectory and remote sensing data facilitates large-scale winter wheat mapping. To the best of our knowledge, this is the first study that fuses operating trajectories of agricultural machinery and satellite images for large-scale crop mapping based on the deep positive and unlabeled learning approach. This study could be possibly applied for better understanding the land cover and land use changes.