Despite advances in AI-driven machine learning or deep learning, challenges persist in precisely evaluating restored mangrove biomass and carbon stocks. Selection of the best potential ML model to the above ground biomass (AGB) is one of major research questions. This research aims at estimating AGB of young mangrove plantation using remote sensing and machine learning. The study leverages three major types of remote sensing data [(i) Sentinel-1 SAR GRD (C-band), (ii) Senitnel-2 MSI L2A (SR), (iii) SRTM DEM data, comprising (401) in total prediction variables], and six feature selection methods [(i) RF-VI, (ii) RF-PI. (iii) PLSR-VIP, (iv) KNN-FIFS, (v) LGBM-PSO and (vi) Higher-order texture features GLCM] as a novel feature optimization approach in this study. In addition, a total of 41 different scenarios as a convergent of multi-source remote sensing data, were explored to identify the optimal feature set for improving model performance. Furthermore, a comparative analysis of nine Machine Learning or Deep Learning Models [RF, SVM, KNN, MLE, XGB-GA, ResNet-18, CNN, MLPNN, & BRNN], was conducted for investigating the best-fitted model in AGB estimation. As a result, ResNet-18 substantially outperforms classical ML. It shows ResNet-18 achieved superior performance (RMSE = 17.612; R2 = 0.772, MAE = 14.696) using scenario_24's optimized multi-sensor features, while conventional ML models (SVM, KNN, RF) showed more modest results (RMSE = 26-28; R2 = 0.40-0.50) in other scenarios, underscoring the critical importance of scenario-specific feature selection beyond model architecture alone. Key predictive features include VH-polarized SAR textures (particularly S1_VH_imcorr1), S2 spectral indices (B3_prom for photosynthetic activity, B11_dent for moisture variation), and GLCM-derived texture patterns, revealing the complementary value of SAR's structural characterization and optical data's spectral information across various modeling approaches. However, some critical gaps remain in mangrove biomass estimation, including the exclusion of below-ground biomass & soil organic carbon, temporal response of the bestfitted models, limited detection of young plantations due to sparse canopies, insufficient species-specific modeling, and limited ground truth dataset.