Selecting Optimal Random Forest Predictive Models: A Case Study on Predicting the Spatial Distribution of Seabed...

Area of Science:

• Marine geology
• Environmental science
• Geospatial analysis

Background:

• Accurate seabed hardness data is vital for sustainable marine management in Australia.
• Current methods for inferring seabed hardness, such as multibeam backscatter and underwater video, have limitations in accuracy and spatial coverage.

Purpose of the Study:

• To develop and evaluate optimal predictive models for classifying seabed hardness into four categories using new schemes (hard90 and hard70).
• To assess the effectiveness of various feature selection methods in improving the accuracy of random forest models for seabed hardness prediction.

Main Methods:

• Developed random forest (RF) models using point data of seabed hardness classes and spatially continuous multibeam data.
• Tested five feature selection (FS) methods: variable importance (VI), averaged variable importance (AVI), knowledge informed AVI (KIAVI), Boruta, and regularized RF (RRF).
• Examined the impact of correlated predictors and identified important/unimportant variables on RF model accuracy.

Main Results:

• The hard90 and hard70 classification schemes effectively categorize seabed hardness.
• Seabed hardness can be predicted with high accuracy using the developed RF models.
• Feature selection methods, particularly AVI and Boruta, identify the most accurate predictive models, challenging the need to exclude highly correlated variables.
• Random forest models demonstrate high predictive accuracy for multi-level categorical data in environmental science.

Conclusions:

• The study validates the effectiveness of the hard90 and hard70 schemes and the high accuracy of RF models for seabed hardness prediction.
• Re-evaluation of variable pre-selection methods is recommended, favoring feature selection techniques like AVI and Boruta.
• Further development of automated computational programs for AVI is suggested to enhance efficiency.