89th American Meteorological Society Annual Meeting

Monday, 12 January 2009: 5:15 PM
Comparison of random forest, artificial neural network, and multi-linear regression: a water temperature prediction case
Room 125A (Phoenix Convention Center)
Robyn L. Ball, Texas A&M University, Corpus Christi, TX; and P. Tissot, B. Zimmer, and B. Sterba-Boatwright
Specific artificial intelligence algorithms are increasingly applied to the modeling of environmental systems. These models are often compared to multi-linear regression models or other reference techniques. Less frequently, artificial intelligence techniques are compared with each other. In this study, two machine learning techniques, an artificial neural network (ANN) and a random forest (RF), are compared to each other and to a multi-linear regression (MLR) model. The features and applicability of the respective models are compared as well as their ability to accurately predict water temperatures in the Upper Laguna Madre, particularly during cold events. Selection of the specific RF model for this case is discussed.

In the Laguna Madre, the longest hypersaline lagoon in the United States, the passage of cold fronts can lower air temperatures by more than 10°C in less than 24 hours which leads to a considerable decrease in water temperature. Records from the past 20 years reveal that some of these cold water events resulted in massive fish kills. In 1997, more than 94,000 fish died in the Lower Laguna Madre and over 48,000 fish died in the Upper Laguna Madre. To mitigate the impact of these cold events, local agencies and stakeholders are considering interrupting activities such as fishing and boating during these events. To help manage such interruptions accurate predictions of occurrences and length of cold water events are needed.

The water temperature predictive models were trained over one year and tested over three years of hourly measurements. Their performance was assessed for varying prediction times (3, 12, 24, and 48 hours). For the 24 hour prediction, the models perform similarly over the test data with an average absolute error of 0.64°C for both the ANN and MLR models and 0.66°C for the RF model. To assess model performance more specifically during cold events, performance was also measured during times when the target water temperature was below 7.2°C. During such cold events, the ANN and MLR models have an average absolute error of 0.50°C and the RF has an average absolute error of 0.61°C. These results confirm the applicability of the models for cold events. In addition, a non-probabilistic contingency table was used to measure the ability of the models to accurately predict the occurrence of cold events. For the 24 hour prediction, the ANN and RF models under-predict by approximately 15% while the MLR model has a small bias of 2%. The true skill statistics of the ANN and MLR models are greater than 0.75 and the RF model's true skill statistic is 0.63 for the 24 hour prediction. Overall performance for shorter prediction times is better while performance dropped for the longer 48 hour predictions.

Although in most cases the ANN and MLR models outperform the RF by a small margin, the RF technique has attractive features that the other models lack. The RF can effectively handle missing data during both the training and testing phases. Because of its ensemble design, it can return a prediction even when some of the input values are missing, making it an appealing model for real-time predictions. Additionally, through the modeling process, a measure of variable importance can be computed, lending insight to the particular system being modeled.

Supplementary URL: