Qing Li, Richard David Williams, Trevor B Hoey, Brian Barrett, Richard Boothroyd

Rivers in the tropics are more likely to exhibit seasonal changes in pattern than those in temperate regions because of strongly seasonal rainfall. However, such changes in seasonal tropical river patterns have not been widely investigated. Machine learning methods are used in this study with Sentinel-2 multispectral remote sensing images to classify active channel landforms (water; unvegetated bars; vegetated bars) of the Bislak, Laoag and Abra Rivers, north-west Luzon, the Philippines. River patterns are classified five or six times per year from 2016 to 2020. Spatial and temporal trends were investigated, in the context of the rivers’ active width, valley confinement, tectonic setting and precipitation. Results show a variety of relationships between each landform unit and active width, but a strong correlation was shown between active width and vegetation area in dry and wet seasons. Rivers were divided into sub-reaches based on observed patterns of water frequency and confinement; Ensemble Empirical Mode Decomposition (EEMD) was then used to decompose the landform time series and precipitation record. EEMD indicates that water and vegetated bars commonly show synchronised fluctuations with precipitation, while unvegetated bars have an anti-phase oscillation with precipitation. It also suggests that deviations from periodic consistency in river pattern may reflect the influence of extreme events and/or human disturbance. At the river system scale, faults perpendicular to the channel centreline were associated with an increase in vegetated bar stability. Overall, the interplay of faults, elevation, confinement and tributary locations impact landform stability. This investigation demonstrates that in tropical regions river pattern should be considered as a dynamic entity as characterising pattern from a single time period may misrepresent a river’s character. EEMD is also demonstrated to be an appropriate statistical technique in geomorphology to decompose datasets that are generated from contemporary applications of machine learning to remotely sensed imagery.