GCFC: Graph Convolutional Fusion CNN Network for Cross-Domain Zero-Shot Extraction of Winter Wheat Map

Abstract

Accurate extraction of winter wheat and its planted area holds significance for agricultural research and government real-time food monitoring. Traditional machine learning methods often demand extensive data and corresponding labels for training in cross-domain classification problems. The heterogeneity of land cover types causes an uneven distribution of samples, leading to unsatisfactory results when these methods are applied directly to other regions. This article introduces a two-branch Graph Convolutional Fusion CNN network incorporating dynamic weighted stratified loss to address these challenges. To reduce the weight of losses generated by easily classified pixels, this loss function adds task masks and category dynamic weights to the cross-entropy loss. Dual branching merges global insights from graph convolutional networks with local emphasis from convolutional neural networks. It enhances the handling of cross-domain classification problems. The first branch introduces an adaptive mechanism and applies it to the graph's adjacency matrix to enhance the model's adaptability to different domain graph structures. The second branch alleviates the oversmoothing problem of edge clustering caused by graph convolution and handles multiscale and spectral information more efficiently. The experimental results showed that the proposed method achieved 99.98% accuracy and good classification results on the Zhoukou dataset. The zero-shot cross-domain prediction on the Suixian dataset achieved 96.11% accuracy. Ultimately, the entire winter wheat planting area of Shangqiu City was extracted with an accuracy of 91.92%. Numerous experiments and practical applications confirm that the proposed method is feasible and effective for winter wheat cross-domain extraction.

Keywords

• Cross-domain
• graph convolution
• oversmoothing
• winter wheat