Coastal marshes are complex and biologically rich ecosystems supporting neighbouring assemblages of aquatic and semi-terrestrial species. They play a critical role in carbon sequestration (CS) and sediment trapping, thus operating as natural filters that improve water quality, supporting neighbouring ecosystems. However, coastal marshes are ecologically sensitive and historically have been threatened by varying factors including man-made infrastructures, global climate change, and meteorological events leading to severe marsh regression. This study presents a fine-grained identification of salt marshes of Venice Lagoon using multispectral Sentinel-2 satellite imagery. The lagoon of Venice is among the largest coastal wetlands in the world that has faced dramatic degradation throughout the last centuries due to the construction of the jetties. This necessitates the task of monitoring the long-term regression and accretion patterns of Venice for researchers to take ecological action against the regression of this large marsh ecosystem. This study presents a compelling case-study of two fine-grained marsh identification methods leveraging micro and macro-spatial context sizes, where micro-spatial inputs act as a magnifying tool for satellite imagery prioritizing entirely on spectral signature deviations, while macro-spatial inputs broaden the receptive context. We comprehensively evaluated the two marsh identification methodologies through holistic, granular, and Explainable Artificial Intelligence (XAI) analysis, thus addressing the black box nature of CNNs. We modified the architecture of various models by employing attention mechanisms resulting in greater performance metrics. We uncovered model-specific gradient-weighted spectral attention relative to the signatures of marsh bodies using Weighted Kernel Density Estimation (WKDE). We compared the correlation between model-specific Spectral Input Importance using Gradient-based techniques with Permutation Analysis by perturbing individual spectral inputs. Finally, we utilized the best performing model reported by our study to quantify decadal marsh and tidal dynamics by monitoring vegetation regression and accretion trends, seasonal marsh floodings, health indices, and chl-a concentration levels.