Spatiotemporal analysis of mangroves using median composites and convolutional neural network

Mangroves play an important ecological role, but these are commonly misunderstood and undervalued. Climatic change and increase in sea level cause risk to these ecosystems. Hence, tracking the extent of mangroves leads to proper management and restoring the damaged ones. Existing index methods for mapping the mangrove extent often struggle with cloud cover, while the studies that feed direct image inputs to models face spatial resolution misalignment, limiting accurate analysis. To overcome these challenges this paper proposes a novel cloud-masked feature extraction (CMFE) approach, integrating bit masking, median compositing, and multi-band spectral sampling for precise mangrove analysis. Our approach utilizes multispectral imagery from United States Geological Survey (USGS) for Landsat 8 and Copernicus for Sentinel 2, leveraging quality assurance (QA) bands for cloud bit masking to nullify cloud and shadow pixels. Cloud cover obscures large portions of satellite imagery and reduces data availability. The median composite technique reconstructs missing or obscured information by aggregating temporal pixel statistics. Spectral and textural attributes are systematically extracted from multiple bands and structured into feature vectors, providing high-dimensional input, which overcomes the spatial resolution discrepancies. Extracted attributes are then fed to a deep learning-based model which classifies the mangroves into different density classes - sparse, moderate, and dense for mapping the dynamic changes over a decade from 2014 to 2024, focusing on Pichavaram Mangrove Forest, Tamil Nadu. The CMFE approach outperforms existing techniques, offering improved accuracy and resilience against cloud cover and spatial misalignment.