Day 4

Detailed paper information

Back to list

Paper title Integrating process-based information into ANN for root zone soil moisture prediction
Authors
  1. Roiya Souissi Université Toulouse III - Paul Sabatier Speaker
  2. Ahmad Al Bitar CESBIO
  3. Mehrez Zribi CNRS
Form of presentation Poster
Topics
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text The estimation of Root-Zone Soil Moisture (RZSM) is important for meteorological, hydrological and mainly agricultural applications. For instance, RZSM constitutes the main reservoir for the crops. Moreover, the knowledge of this soil moisture component is crucial for the study of geophysical processes such as water infiltration and evaporation. Remote sensing techniques, namely active and passive microwave, can retrieve surface soil moisture (SSM). However, no current spaceborne sensor can directly measure RZSM because of their shallow penetration depth. Proxy observations like water storage change or vegetation stress can help retrieve spatial maps of RZSM. Land surface models (LSM) and data assimilation techniques can be also used to estimate RZSM. In addition to these methods, data-driven methods have been widely used in hydrology and precisely in RZSM prediction. In a previous study (Souissi et al. 2020), we demonstrated that Artificial Neural Networks (ANN) can be used to derive RZSM from SSM solely. But we also found limitations in very dry regions where there is a disconnection between surface and root zone because of high evaporation rates.
In this study, we investigated the use of surface soil moisture and process-based features in the context of ANN to predict RZSM. The infiltration process was taken into account as a feature through the use of the recursive exponential filter and its soil water index (SWI). The recursive exponential filter formulation has been widely used to derive root zone soil moisture from surface soil moisture as an approximation of a land surface model. Here, we use it only to derive an input feature to the ANN.
As for the evaporation process, we integrated a remote sensing-based evaporative efficiency variable in the ANN model. A very popular formulation of this variable, defined as the ratio of actual to potential soil evaporation, was introduced in (Noilhan and Planton, 1989) and (Lee and Pielke, 1992). We based our work on a new analytical expression, suggested for instance in (Merlin et al., 2010), and replaced potential evaporation by potential evapotranspiration that we extracted from the Moderate Resolution Imaging Spectroradiometer (MODIS) Evapotranspiration/Latent Heat Flux product.
The vegetation dynamics were considered through the use of remotely sensed Normalized Difference Vegetation Index (NDVI) from MODIS.
In-situ surface soil temperature, provided by the International Soil Moisture Network (ISMN), was used. Different ANN models were developed to assess, each, the impact of the use of a certain process-based feature in addition to SSM information. The training soil moisture data is provided by the ISMN and is distributed over several areas of the globe of different soil and climate parameters. An additional test was conducted using soil moisture sensors not integrated to the ISMN database, over the Kairouan Plain which is a semi-arid region in central Tunisia covering an area of more than 3000 km2 and part of the Merguellil watershed.
The results show that the RZSM prediction accuracy increases in specific climate conditions depending on the used process-based features. For instance, in arid areas where ‘Bwh’ climate class (arid desert hot) is prevailing like eastern and western sides of the USA and bare areas of Africa, the most informative feature is evaporative efficiency. In areas of continental Europe and around the Mediterranean Basin where there are agricultural fields, NDVI is for example the most relevant indicator for RZSM estimation.
The best predictive capacity is given by the ANN model where surface soil moisture, NDVI, recursive exponential filter and evaporative efficiency are combined. 61.68% of the ISMN test stations undergo an increase in correlation values with this model compared to the model using only SSM as inputs. The performance improvement can be also highlighted through the example of the Tunisian sites (five stations). For instance, the mean correlation of the predicted RZSM based on SSM only strongly increases from 0.44 to 0.8 when process-based are integrated into the ANN model in addition to SSM.
The ability of the developed model to predict RZSM over larger areas will be assessed in the future.