Pattern recognition describing spatio-temporal drivers of catchment classification for water quality

Area of Science:

• Hydrology and Water Quality
• Environmental Science
• Artificial Intelligence in Environmental Modeling

Background:

• Spatial data classification is crucial for hydrological modeling, especially in ungauged regions.
• Current land use classifications yield inconsistent water quality results, hindering effective decision-making.
• The influence of temporal variations (season, flow) on classification consistency and the role of biotic responses remain underexplored.

Purpose of the Study:

• To investigate if temporal variations in water quality drivers affect classification consistency.
• To determine if spatial datasets incorporating original vegetation can capture biotic response variability.
• To enhance the accuracy of hydrological models for ungauged catchments through improved spatial classification.

Main Methods:

• Utilized Artificial Neural Network Pattern Recognition (ANN-PR) to match catchments based on Dissolved Inorganic Nitrogen (DIN) patterns.
• Partitioned water quality datasets by season (Wet vs. Dry) and flow (Increasing vs. Retreating).
• Employed explainable AI for catchment classification using spatial feature datasets and validated findings with Kruskal Wallis tests.

Main Results:

• Seasonal partitioning of water quality data yielded the highest corroboration rates between spatial and DIN classifications.
• Classified catchments showed significant independence (p < 0.001 to 0.026) when using partitioned, rather than non-partitioned, datasets.
• Identified three Dissolved Inorganic Nitrogen (DIN) pattern categories, with vegetation types (woodlands, vineforest) acting as key indicators for classification under different temporal scales.

Conclusions:

• Temporal variations, particularly seasonal changes, significantly improve the consistency of land use classification for hydrological modeling.
• Incorporating original vegetation data as a proxy for biotic responses enhances the reliability of spatial data for classifying ungauged catchments.
• This approach offers a pathway to more accurate water quality predictions and better-informed data strategies for hydrological research.