Clustering relies on a single measure of "distance" between objects, so it's just a matter of devising a reasonable way combining several distances (one for image data, one for physical measurements, one for pollinators, ...) into one. It is subjective, and that is both an advantage and a disadvantage. It's an advantage because you can emphasize the contributions that you feel are most important for your analysis, and it's a disadvantage because there is no one right way to do it.

I mentioned the Fisher iris dataset because there are 50 examples for each of the three species, and there is a fair amount of variability of the four measurements within each species.

An analysis that might be useful with pollinator data is formal concept analysis (FCA, see https://link.springer.com/book/10.1007/978-3-642-59830-2). In this case the objects are the different species of orchid and the attributes are the pollinator species that attend each orchid species. You could also include habitat (soil type for terrestrial and host species for epiphytic) among the attributes, and there is even a way of scaling quantitative measures. It would take some creativity to make use of image data in FCA, but I don't think it's impossible.

You should be able to get useful results from both methods by including more examples of each species or cultivar, especially varying the backgrounds.

You should experiment with different image feature extractors, different distance functions, and different cluster dissimilarity functions to find the combination that gives you results that you find useful. You might also try training a classifier with Classify (again experimenting with these same options), and then test it on images of the same species that were not in the training images.

Are you familiar with Fisher's iris data set?

In[4]:= ExampleData[{"MachineLearning", "FisherIris"}, "Properties"]

Out[4]= {"Data", "Description", "Data", "Dimensions", "LearningTask", \
"LongDescription", "MissingData", "Name", "Source", "TestData", \
"TrainingData", "VariableDescriptions", "VariableTypes"}

While the data are not images, they are often used to demonstrate the characteristics of different clustering methods.

Please post your code, especially for the second example (it looks like you tried using one of the examples from the Applications section of ImageDistance). You cloud use ImageDistance to form a distance matrix from your collection of images, and then use that as input to ClusteringTree.

What is "EMD"? I couldn't find it anywhere of the ImageDistance reference page.