Enhancing the predictive performance of remote sensing for ecological variables of tidal flats using encoded features from a deep learning model