From the article you cite, I have found this figure which looks similar to yours.
On the same page of the article, they say:
Having a smooth profile also plays a role in better gradient flow, as
shown in Fig. 3, where the output landscapes of a five-layered randomly
initialized neural network with ReLUand Mish are visualized. The
landscapes were generated by passing in the co-ordinates to a
five-layered randomly initialized neural network which outputs the
corresponding scalar magnitude. The output landscape of ReLU has a
lot of sharp transitions as compared to the smooth profile of the
output landscape of Mish.
It seems your figure (and the one above) does not show the loss landscapes (which is the figure 4 in the article) but the landscape of the global neural network using different activation functions. I guess the neural network they used to produce these figures has 2 inputs (the coordinates of the plots), 3 hidden layers and one scalar output. The color in the plots corresponds to the scalar output value.
The authors present these figures because :
Smoother output landscapes suggest smooth loss landscape [...]
This is my understanding of the article, hope it helps.