Mapping Winter Wheat Using Ensemble-Based Positive Unlabeled Learning Approach

High-resolution remote sensing images can support machine learning methods to achieve remarkable results in agricultural monitoring. However, traditional supervised learning methods require pre-labeled training data and are unsuitable for non-fully labeled areas. Positive and Unlabeled Learning (PUL), can deal with unlabeled data. A loss function PU-Loss was proposed in this study to directly optimize the PUL evaluation metric and to address the data imbalance problem caused by unlabeled positive samples. Moreover, a hybrid normalization module Batch-Instance-Layer Normalization was proposed to perform multiple normalization methods based on the resolution size and to improve the model performance further. A real-world positive and unlabeled winter wheat data set was used to evaluate the proposed method, which outperformed widely used models such as U-Net, DeepLabv3+, and DA-Net. The results demonstrated the potential of PUL for winter wheat identification in remote sensing images.