Understanding limitations in generalizability and performance across two process guided deep learning architectures for predicting stream temperature
Abstract
Water temperature is a fundamental indicator of ecosystem health that contributes directly to biological metabolism, nutrient cycling, and habitat suitability. Globally, anthropogenic stressors such as dams and urbanization are altering the water temperature regimes of streams and rivers. These alterations are expected to be exacerbated by climate change through both rising temperatures and changes in streamflow. Existing process-based, statistical, and machine learning models have made significant progress in estimating water temperature in both monitored and unmonitored locations. However, the uncertainty for these models remains at ecologically significant magnitudes, making it difficult to use them to predict changes in ecosystem function. Recent efforts that combine physical models with deep learning have shown significant promise in improving the predictive capability of water temperature models. While these efforts, referred to as process guided deep learning, have been shown to outperform stand alone process based and purely machine learning models, there is still significant uncertainty in how different deep learning architectures perform in regards to input biases, generalizability to out-of-bound predictions, and overall computational efficiency and performance. Here, we compare two spatiotemporally aware process guided deep learning architectures to better understand their benefits and limitations when predicting stream temperature. The first is a recurrent graph convolution model driven by a long-short term memory network, while the second relies on gated temporal convolution layers. While both models have high predictive capability (RMSE of 1.73 and 1.75 °C respectively), they vary in their computational efficiency and ability to handle systematic biases in training data. Understanding these limitations is essential moving forward as we try to predict future stream temperatures under changing climate and precipitation regimes.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H33J..10T