Abstract: Given the similar spectral characteristics of mangroves and Spartina alterniflora during their growing season, the mangrove identification faces challenges, such as mixed classification or omission. In addition, some scattered small-scale mangroves are often ignored or misclassified in land cover statistics. To solve the above problems, this study selected appropriate mangrove sensitivity indexes based on multi-source remote sensing data from Landsat8, Landsat9, and Sentinel-2 satellites, along with neural network models. Then, combined with spectral characteristics, phenological characteristics, and spectral-temporal features, this study achieved accurate extraction of small-scale mangroves. The results indicate that the neural network-based classification model achieved the land cover classification of vegetation, bare flats, and water bodies, with an overall classification accuracy of 97.59%. However, its performance in mangrove extraction was limited, with an accuracy of only 93.57%. By combining phenological characteristics and spectral-temporal features, the accuracy of mangrove extraction was improved to 96.24%. Applying the model in the intertidal zones of Yueqing Bay, the identification results show, between 2018 and 2023, a significant expansion in mangrove distribution range (average: 25.62 hm²/a) and large-scale manual removal of Spartina alterniflora (about 2 720.39 hm²). The Yueqing Bay exhibited an overall trend of land transfer, involving the conversion of Spartina alterniflora and bare flats to mangroves, as well as the conversion of some Spartina alterniflora to bare flats. The study provides a methodological framework for the extraction and phenological analysis of small-scale mangroves. The proposed extraction method has wide applicability and can provide data support for the ecological protection and restoration of mangroves.