Optimising Irish Fen Habitat Mapping: An Evaluation of Temporal Windowing and Spectral Features Using Google Earth Engine and Random Forest

High spectral similarity and overlapping habitat boundaries present a substantial challenge for mapping and monitoring Ireland’s ecologically vital fen ecosystems. This study establishes a reproducible remote sensing workflow by combining Google Earth Engine (GEE) with Sentinel-2 imagery for Irish fen classification. This study uses a Random Forest machine learning model to evaluate five temporal windows against three spectral feature sets across three distinct sites. The results indicate that a growing season window (June–September) incorporating both bands and spectral indices captures peak phenological information more effectively than annual composites, producing high within-site, pixel-level overall accuracy across all study sites (Scragh Bog: 95.24%, Liskeenan Fen: 94.13%, Ballymore Fen: 94.43%). Although raw spectral bands achieved marginally higher localised accuracies at Liskeenan (+0.24%) and Ballymore (+1.65%), the combined feature set was maintained as the standard workflow to ensure cross-site stability, biophysical parameterisation, and resilience against localised soil background noise. To accommodate narrow, fragmented geometries, training data was extracted via pixel-level stratified sampling; while this introduces spatial autocorrelation, classification outputs showed coherent within-site spatial patterns, although accuracy estimates remain subject to spatial autocorrelation. Crucially, the Google Earth Engine-based Random Forest model maintained stable within-site classification performance when reducing the training-testing ratio from a standard 80:20 split down to a highly constrained 50:50 split. Ultimately, this workflow provides a cloud-based framework for monitoring vulnerable wetland ecosystems under real-world geometric and ground-truth data limitations.