Simultaneous emulation and downscaling of modeled soil state variables with machine learning
Description
We propose a lightweight, computationally efficient machine learning (ML) model capable of emulating the LIS-based soil moisture and soil temperature and downscaling them from a native 10 km resolution to 1 km resolution. Our approach is extendable to other variables as long as a non-linear relationship between meteorological forcing and the variable of interest can be conceptualized as modulated by local conditions (elevation, soil type, land cover, vegetation). Then, a branched neural network (NN) architecture structurally represents this relationship. As a part of the project, different NN architectures and input combinations have been tested and assessed using SHapley Additive exPlanations (SHAP) values and ablation analysis. Currently, the downscaled product is being validated and compared to other high-resolution products.
Detailed example
The model outputs are emulated LIS-like soil moisture and soil temperature (predictions), as well as downscaled soil moisture and soil temperature.
AI / analytics pattern
Classical/Predictive Machine Learning: Models trained on data to make predictions or classifications based on identified patterns or relationships.
Automation level / stage
a) Pre-deployment – The use case is in a development or acquisition status.
Expected benefit
Using the proposed method, it is possible to obtain LIS-like quality predictions for soil state variables in seconds, as well as provide unprecedented for LIS downscaled to 1 km data relevant to a wide variety of applications. The low computational cost of the inference and the ability to resolve fine-resolution features expedite obtaining the crucial information for decision-making.
Controls / human review
ATO: Not reported; PIA: Not published