Figure 6
Temporal and spatial covariance among depth change, bower construction, feeding, and spawning. Hierarchical clustering of depth change and 3D ResNet-predicted behavioral events through time reveals three largely distinct behavioral clusters corresponding to feeding (left-most columns), bower construction (right-most columns), and quivering behaviors (middle column, (A); Pearson’s R values shown for each pairwise correlation). Quivering was more strongly and positively correlated with bower construction than with feeding. A similar analysis of the spatial distributions of depth change and behavioral events across the sand surface revealed two major clusters, one corresponding to feeding (left-most columns), and the other corresponding to quivering and construction behaviors (right-most columns; (B)) notably, quivering was strongly spatially correlated but weakly temporally correlated with bower construction behaviors. To analyze social context, we used object recognition to detect and count fish above the sand tray during different behaviors, with green outlines indicating predicted fish (C–F); 1, 2, 3, and 4 fish detected, respectively). Analysis of the number of fish present above the sand tray during 3D ResNet-predicted behavioral events revealed strong differences in fish count across behavioral categories ((G); p < 2.2 × 10–16). Fish counts were lowest during bower construction, greater during feeding, and greatest during spawning, supporting dynamic and interrelated behavioral and social contexts across whole trials. Pearson’s R values are shown for each pairwise comparison, and the significance of each pairwise comparison and corresponding p-values can be found in Supplementary Tables S1 and S2. Letter abbreviations for each behavioral measure are indicated in the labels along the horizontal axis.
