Mapping the invasive Spartina alterniflora in sub-meter level with improved phenological spectral features and deep learning method

Abstract

The invasion of Spartina alterniflora (S. alterniflora) has severely impacted China’s coastal ecosystems by disrupting native habitats and polluting waters, highlighting the necessity for accurate distribution mapping. Existing studies mainly rely on moderate-resolution remote sensing imagery, like Landsat and Sentinel-2, but caused by the limitations of spatial resolution, mixed pixels have led to the misclassification of small patches and boundaries. This study developed sub-meter phenological spectral features of S. alterniflora and improved the DeepLabv3+ model with a Class Feature Attention Mechanism (CFAM) to produce the first sub-meter S. alterniflora product for the Beibu Gulf of China in 2020-2021. The results indicated that the developed sub-meter spectral features and the improved DeepLabv3+ model could enhance classification performance. The total area of the sub-meter product in this study was 1,190.36 hectares, which was 83.17 hectares less than the 10-meter product. When benchmarked against the sub-meter product, the 10-meter product exhibited an omission of 314.36 hectares and a commission of 397.53 hectares, with a spatial discrepancy of 711.89 hectares. This method provides a new approach for fine-scale invasive species monitoring. The sub-meter S. alterniflora distribution product provides critical baseline data for monitoring and managing S. alterniflora in the Beibu Gulf.

Keywords

• Spartina alterniflora
• Invasive species
• Sub-meter mapping
• Phenology
• Deep learning