Performance of GEDI data combined with Sentinel-2 images for automatic labelling of wall-to-wall corn mapping

Abstract

Corn is the dominant crop planted in Northeast China, and its accurate and timely mapping is important for food security and agricultural management in China. However, the absence of enough labels is challenging for corn accurate mapping in a regional area using machine learning methods or deep learning methods. In this study, an efficient way of automatic labelling and mapping of corn planted areas by combining Global Ecosystem Dynamics Investigation (GEDI) data and Sentinel-2 images is proposed. We explore the height and vertical structure differences between corn and other crops derived from GEDI features and generate labels automatically by referencing crop type products and transferring models from historical years. The trained learning networks of automatic labelling from GEDI points and the trained decision trees of the Random Forest (RF) classifier can be transferred to corn mapping in arbitrary target years. The Sentinel-2 features are combined to perform wall-to-wall corn mapping using a random forest algorithm and GEDI-based labels. This approach is used to map corn planted areas in Northeast China from 2019 to 2022, and the classification results are validated using independent labels collected in field campaigns in 2023, published maps, and official statistics. Our classification results reveal that our proposed method achieves high accuracy and robustness with an average overall accuracy of 0.91 validated using testing labels from spatial-type stratified sampling. The correlation coefficient (R2) between our classified result with the official statistical data and published classification results reach 0.96 and 0.98, respectively. These results demonstrate the potential of GEDI data for automatic label collection for vegetation with height difference and provide a new approach for efficient crop mapping on a large-scale.

Keywords

• GEDI
• Automatic labelling
• Wall-to-wall mapping
• Model transfer
• Google Earth Engine