Since it is a number I thought of a Poisson distribution for the dependent variable. Thus a linear model with forest and frogs as explanatory factors. That can be estimated with a generalised linear model. I believe that they have hat in spss. Still the number is very low so, I am sorry, but I guess that it will be difficult to get precise estimates.

Don't use the Tukey HSD test or Duncans test. Duncans test is from the 1950:ies when they had no clear view about multiple inference, so the error rate is larger than the promised 5%. Tukeys hsd test (is great otherwise) but it is based on the studentized range distribution, which is based on the normal distribution, and these data are not normal. Besides when comparing among only three forest it would have been better to use the Bonferroni method, or preferably Bonferroni-Holm. But, the sample size is so small anyway so I suggest to skip all the multiple comparisons anyway.

Besides, I start to wonder: Are you infering about forests in general or was it just three forest that were investigated? (I would guess so.) You could possibly have sampled from ten times three different forests. Now it seems like the conclusion can only be about these three forests (an not forests in general).

Also about the choice of frogs species. The usual thing would be to choose a fixed set of frog species in advance. Now, maybe you have been crawling around in the forests and noted the one you happened to discover. (There is no specie with only zeros.) I am not sure of what would be the statistical consequences be of such a procedure. To be frank, I believe that most would just ignore tricky little things like this, but still the problem can be there.

Let's see if the other ones have any solutions.