Emergency-oriented fine change detection of flood-damaged farmland from medium-resolution remote sensing images

• Effectiveness of fine change detection model in flood-damaged farmland. • Propose a semi-supervised sample generation framework. • Construct a different degrees of flood-damaged farmland dataset. • Verify robustness of SSLSampleGen across regions and remote sensing satellites. Flood disasters are characterized by frequent and sudden occurrences and obvious chain effects, posing a major threat to agricultural production. Government disaster relief and agricultural insurance are increasingly urgent in assessing losses to flood-damaged farmland. There are many challenges in assessing flood-damaged farmland. On the one hand, the historical data needed for the farmland loss assessment model is missing. On the other hand, the farmland loss assessment conducted within the flood inundation area only focuses on the completely submerged farmland, which is not accurate enough. Semi-supervised deep learning can effectively alleviate the above problems. This study specifically designed a fine-grained change detection model for flood-damaged farmland using medium-resolution remote sensing (RS) images to achieve multi-scenario change detection of flood-damaged farmland. In order to assist in the automatic sample generation of large-scale unlabeled sample RS images of flood status, a semi-supervised sample generation framework for flood-damaged farmland using medium-resolution RS images is proposed. Based on this framework, FloodedCropland datasets is created. The experimental results show that the proposed change detection model has an F1-score of 0.9047 on flood-damaged farmland. After the semi-supervised sample generation framework optimized the model, the change detection F1-score was improved to 0.9241. Experiments have verified that the automatic generation of labels for flood-damaged farmland in medium-resolution RS images using a semi-supervised sample generation framework performs better than the scarce manual labeling model and can save a lot of manual labeling time. The consistent performance in different geographic regions and under different RS satellites imaging conditions demonstrate the practical application potential of this method for cross-regional and cross-RS satellites intelligent information extraction in natural disaster scenes with a lack of labeled samples.