A Generative Time-Series Data Augmentation Framework Based on Multisource Remote Sensing for Early Crop Classification

Accurate and timely crop spatial classification based on remote sensing imagery is a crucial and challenging task in the agricultural field. In the early stages of crop growth, the available spectral features of crops in the temporal dimension are limited. Besides, the spectral curves of crops such as soybean and corn are highly similar during this period. It is only after the key phenological stages (flowering stage for corn and podding stage for soybean) that accurate differentiation between these two crops can be achieved. To address the issue of insufficient effective features and missing key characteristics in early classification tasks, this study proposes a time-series prediction and generative data augmentation framework based on multi source remote sensing for early crop classification. This method is inspired by traditional approaches for generating discrete remote sensing images and further enables the continuous generation of spectral data along the temporal dimension. We have designed a comprehensive workflow to generate the spectral curve trends of different crops during their early growth stages and generate feature data to support classification tasks. We conducted early classification experiments for rice, soybean, and corn in Northeast China. These findings demonstrate that the method significantly enhances the accuracy, earliness, and stability of early crop classification tasks.