I really like this clustering approach, "over geographical and OTHER dimensions".

Question: How do you think this can be applied to automatically build risk-sharing groups of smallholder farmers (like in India) whose fields share the same satellite vegetation index response to weather hazards?

We are building a system that is currently deployed in Tamil Nadu, India, helping more than 300 farmers. We group them by geographical proximity but it's not ideal. For example fields may be far away but spread along a same river they share the same response.