Addressing hydrological modeling in watersheds under land cover change with deep learning. (August 2021)

• Record Type: Journal Article
• Authors: Althoff, Daniel; Rodrigues, Lineu Neiva; Silva, Demetrius David da
• Is Part Of: Advances in water resources. Volume 154(2021)
• Publication Date: 2021-08
• Subjects: Cerrado -- Data-driven -- Explainable artificial intelligence -- Regional hydrological model; Hydrology -- Periodicals; Hydrodynamics -- Periodicals; Hydraulic engineering -- Periodicals; 551.48
• DOI: 10.1016/j.advwatres.2021.103965
• Languages: English
• ISSNs: 0309-1708

Abstract

Highlights: An LSTM hydrological model is developed considering land cover characteristics as dynamic in time. The integrated gradients technique is used to interpret the importance of features and dependency on lookback series for the model. Previous hydrological knowledge of the cerrado is compared to what the model learned. Abstract: The impacts of land cover change have traditionally been assessed in hydrological modeling with a priori knowledge, e.g., using methods based on the curve number, or by calibrating hydrological models over different time periods. However, how hydrological processes respond to such changes is extremely context-dependent. Thus, there is an opportunity for the development of hydrological models that can learn from large hydrological data sets under the context of severe environmental changes. In this study, a single regional hydrological model is developed based on long short-term memory (LSTM) neural networks using different input configurations. One model considers only meteorological forcings as inputs (I1), another model considers meteorological forcings and static catchment attributes (I2), and a third model also considers meteorological forcings and catchment attributes but where the land cover characteristics are dynamic (I3). The models are trained using information from 411 catchments in the Brazilian Cerrado biome. The data set includes, for each catchment, the daily streamflow observations (target), daily precipitation and reference evapotranspiration (meteorological forcings), and 21 catchment attributes including topography, climate indices, soil characteristics, and land cover characteristics. Considering catchment attributes increases the performance of the LSTM model (I2 and I3 median KGE : 0.69). Considering the land use cover characteristics as dynamic improves the predictions under low-flow conditions (I3 median rNSE : 0.62) when compared to the model considering such characteristics as static (I2 median rNSE : 0.53). This study also uses the deep network with the integrated gradients technique to explore the contribution of the catchment characteristics to streamflow and the number of time steps of influence for the deep network in different regions.