The ensemble learning combined with the pruning model reveals the spectral response mechanism of tidal flat mapping in China

Abstract

Tidal flats play a crucial role in biogeochemical cycles, and the mapping of tidal flats is essential for coastal ecological protection. Remote sensing technology offers a powerful tool for large-scale mapping of tidal flats distribution. However, understanding the spectral response mechanism of tidal flats remains a challenge. This research utilized Rule Combination and Simplification (RuleCOSI+) to automatically prune Random Forest (RF) trees, enabling a more interpretable explanation of the black-box model and uncovering the spectral response mechanisms of tidal flats using Sentinel 1/2 imagery. By simplifying the RF, the number of rules was reduced by 99.7 %, from 11,587 to just 32, with only a 1 % decrease in overall accuracy (from 96.4 % to 95.4 %). Similarly, the identification of muddy and sandy tidal flats has also been simplified, with the number of rules reduced from 2018 to 18, a decrease of 99.1 %, while the accuracy increased by 1.2 % (from 97.4 % to 98.6 %). The simplified rules significantly reduce the complexity of understanding the spectral response mechanisms of tidal flats while enabling flexible and rapid mapping across different regions and periods. The soil moisture content was the dominant factor in tidal flat identification, with vegetation and built-up land indices providing supplementary information to distinguish other land types. Notably, the shortwave infrared response to moisture proved critical for distinguishing between muddy and sandy tidal flats. These findings offer valuable insights into the remote sensing mechanisms underlying tidal flat identification and can serve as a reference for interpreting other land use types or classification systems.