Accurately quantifying the carbon sequestration capacity of coastal blue carbon ecosystems (mangroves, salt marshes, and seagrass beds) is essential for developing nature-based solutions to mitigate global climate change. These ecosystems act as significant carbon sinks, yet traditional field-based assessments are often labor-intensive, time-consuming, and spatially limited. In recent years, the integration of multi-source remote sensing data with machine learning (ML) techniques has gained increasing attention, offering new possibilities for large-scale, high-resolution blue carbon quantification. However, there are limited reviews focusing on machine learning algorithms for quantifying blue carbon and the types of remote sensing data sources that can be used presenting a need for a systematic summary of pathways for applying multi-source data and machine learning algorithms to blue carbon ecosystems. This study systematically reviews the applications of machine learning in coastal blue carbon quantification from the recent five years, focusing on the utilization of classification and prediction frameworks. Utilizing multi-source datasets-including satellite, airborne, and drone-based sensors-these frameworks harness the strengths of various machine learning algorithms. Classification methods, such as Random Forest (RF), Support Vector Machines (SVM), and deep learning models, are primarily employed for mapping vegetation types and assessing spatial distributions. Prediction models, such as XGBoost, CatBoost, and Cubist, estimate carbon stocks as continuous variables by incorporating spectral, structural, and environmental features. Each approach presents distinct advantages-classification excels in large-scale ecosystem mapping, while prediction models provide high-accuracy stock assessments when reliable and sufficient training data is available. Combining these methods in a "classification-first, then prediction" framework enhances both spatial and quantitative precision. Despite significant progress, challenges persist in feature selection, data heterogeneity, and the interpretability of machine learning outputs, limiting model accuracy and applicability. Spectral signature variability, particularly in distinguishing similar vegetation types, and radiative saturation in dense biomass areas like mangrove forests introduce classification errors. The integration of Synthetic Aperture Radar (SAR), LiDAR, and hyperspectral imaging can mitigate these limitations by improving structural and biochemical feature extraction. Additionally, underwater remote sensing technologies such as autonomous underwater vehicles (AUVs) with optical sensors offer promising solutions for monitoring submerged seagrass meadows. Advancements in hybrid machine learning models and explainable artificial intelligence (AI) techniques can further enhance model reliability, ensuring more accurate and interpretable blue carbon quantification. As AI and remote sensing continue to evolve, their synergy presents new opportunities for refining carbon accounting methodologies and informing climate policy. Strengthening interdisciplinary collaboration between ecologists, data scientists, and policymakers will accelerate progress in this field, ensuring that machine learning-driven approaches contribute meaningfully to global carbon neutrality efforts.