TL;DR
This paper explores using deep learning (specifically Feed-Forward Neural Networks and Gated Recurrent Unit networks) to create "emulators" that mimic complex agricultural simulation models like APSIM. The goal is to speed up "what-if" scenario analysis for environmental management, but they found that opportunistic use of existing datasets has limitations in emulating all output variables accurately.
What Problem Does This Paper Address?
Environmental models, like the Agricultural Production Systems sIMulator (APSIM), are crucial for understanding and managing complex systems such as agricultural catchments and water quality. However, these models are computationally expensive, requiring millions of simulation runs for comprehensive "what-if" scenario analysis. This can take months, limiting the scope and speed of research and policy-making. Model emulation offers a solution by creating computationally efficient surrogate models, but building effective emulators often requires carefully designed datasets that may not always be available. This paper investigates the feasibility of using existing, "opportunistic" datasets (not specifically designed for emulation) to train deep learning models for APSIM emulation.
Methods
The researchers employed two main deep learning architectures:
- Feed-Forward Neural Networks (FFNNs): Standard neural networks where information flows in one direction through layers.
- Gated Recurrent Unit feeding into FFNN (GRU-FFNN): A type of recurrent neural network (RNN) that uses GRUs to process sequential data (like daily time steps in a simulation) before feeding into an FFNN. RNNs are generally better at capturing temporal dynamics.
They also implemented a clustering approach based on correlation statistics to group similar APSIM model runs, aiming to create more homogeneous training datasets. Input features were scaled to ensure numerical stability during training.
Datasets
The primary dataset used was an existing, large collection of APSIM model runs generated for the Paddock to Reef (P2R) Integrated Monitoring, Modelling and Reporting Program. This dataset was not specifically tailored for model emulation. It contains time-series data for numerous variables, with each model run simulating decades of daily states.
Key Insights and Ideas
- Opportunistic data has limits: While existing large datasets are valuable for initial exploration, they may not be "fit-for-purpose" for deep learning emulation, especially for complex dynamics.
- Not all outputs are equal: Some simulation outputs (like runoff and dissolved inorganic nitrogen in runoff) are easier to emulate accurately than others (like soil loss and leached nitrogen).
- GRU-FFNNs show promise: The recurrent nature of GRU-FFNNs appears to be more effective for capturing the temporal dependencies needed for accurate emulation compared to standard FFNNs.
- Clustering helps, but isn't a silver bullet: Grouping similar model runs improved emulation, but lowering the correlation threshold to include more data sometimes degraded the emulation of certain variables.
- Data richness is key: The sparsity of data and limited temporal variation in the opportunistic dataset likely hindered the ability of the emulators to learn complex relationships, particularly for soil loss.
Key Results
- The GRU-FFNN architecture with 3 hidden layers and 128 units per layer provided good emulation for runoff and DINrunoff.
- Soil loss and Nleached were emulated relatively poorly across various architectures, failing to capture variability at higher output values.
- Lowering the minimum correlation threshold for clustering improved Nleached emulation but worsened DINrunoff emulation, highlighting a trade-off.
- The paper visually demonstrates emulation performance using scatter plots, showing how well the emulated outputs match the original outputs.
Limitations and Future Work
The authors acknowledge that the opportunistic dataset, despite its size, has limitations:
- Sparsity: Many variables have long periods of zero values.
- Temporal similarities: Meteorology settings (like rain patterns) might be too similar across runs, limiting the learning of diverse dynamics.
- Data richness: The dataset may not be rich enough to capture the complex dynamics of all output variables.
Future work suggested includes:
- Design of Computer Experiments (DoCE): Proactively generating datasets specifically tailored for emulation, ensuring a wider range of inputs and more informative data.
- Synthetic Meteorology: Creating long, synthetic meteorology files to feed through APSIM scenarios, generating a denser and more varied dataset for training.
- Exploring other RNN architectures like LSTMs.
Summary updated