Extended Data Fig. 4: Acoustic seascape predictive model.

A Random Forest learning algorithm was used to predict the acoustic seascape classification from environmental variables. The model was trained with sea surface temperature (a), subsurface dissolved oxygen (b) and chlorophyll (c) as inputs, and with the acoustic seascape classification as the output (d). An example of predicted classification is shown (e) with the average importance of predictors (f) and the rate of success (F1 score, from 0 to 1) to predict the subpolar (SP), Gyre (G), subtropical (ST), tropical (T), upwelling (UW), and low-oxygen (LO) seascape classes (g). Environmental inputs were extracted as monthly-weighted averages from the period 2000–2020 based on the monthly acoustic data coverage at each 4×4 degrees cell (see Supplementary Fig. 1 and Methods). The model accuracy was evaluated by comparing 25% subsets of the observed classification against predictions trained with the remaining 75% of data. The operation was repeated 100 times with random data subsets for testing and training. On average, the prediction accuracy of the model was 0.78 ± 0.04%.