Winter wheat mapping without ground labels via automated knowledge transfer across regions and years

Abstract

Accurate and timely information on winter wheat distribution is essential for agricultural management and food security. However, automated approaches for large-scale winter wheat mapping are often hindered by the scarcity of training data. Moreover, well-trained classification model is usually applicable to specific spatial and temporal scales. This study proposed an automated knowledge transfer approach based on adaptive segmentation of phenological similarity images (KT-SimSeg), to shed the reliance on ground labels and reduce spatiotemporal variability for cross-region/year winter wheat mapping. The optical and radar phenological patterns of winter wheat, constructed by the prior knowledge in source domains, were transferred to target domains and aligned with the temporal sequences of undefined pixels. Two-layer phenological similarity between winter wheat and undefined pixels were calculated as indicators to label undefined pixels automatically based on an adaptive threshold segmentation algorithm. Resulting labeled pixels were further refined and were used to pre-train a random forest (RF) model for winter wheat mapping. The performance of the KT-SimSeg approach was assessed in seven regions across the globe for three years, and further compared with two model transfer approaches based on RF (MT-RF) and one-class support vector machine (MT-OCSVM) classifier. The proposed approach performed well for winter wheat mapping with F1-score values of 0.925 and 0.914 across regions and years, which was superior to either MT-RF (0.733 and 0.913) or MT-OCSVM (0.383 and 0.836). Besides the well-delineated winter wheat parcels, the planting areas detected by the KT-SimSeg also showed strong correlations (R² = 0.93 and 0.94) with the reference across spatial and temporal domains. Additionally, the KT-SimSeg could identify winter wheat accurately as early as the heading stage. The proposed approach offers a viable solution to produce high-quality regional winter wheat products without local ground labels, and has potential for knowledge transfer across crop production regions and years in quantitative remote sensing modeling.