LACC: A lightweight attention-conditional convolution network for long-term Wetland classification

Wetlands are critical in keeping ecological balance and supporting biodiversity. However, wetland degradation has become more severe. It is crucial to understand the ecological and biological conditions within them to address these threats. Recent advances in artificial intelligence give new tools for wetland monitoring. The study uses Landsat because it is well-suited for long-term and regional-scale analysis for its resolution and continuous data. A novel framework-the Lightweight Attention-Conditional Convolution (LACC) network-is introduced to efficiently process extended time series. By combining lightweight attention with conditional convolution, LACC is designed to better capture temporal variability and complex class patterns while remaining computationally efficient for large-area applications. Using the new model, we generated a 20-year wetland land-cover dataset for the whole state of Louisiana, providing valuable insights into the wetland conditions. The time series could be used to link class transitions and vegetation conditions to changes in biomass/soil carbon stocks, as well as methane-emission proxies, thereby strengthening regional GHG inventories and connecting wetland monitoring to carbon accounting and urban resilience planning across coastal Louisiana.