Something our team has learnt over the last 4 years is the geographical scope usually plays a big role in AI vision models (detection, segmentation and classification tasks), local variability is huge in some regions (e.g the tropics). Thus the usefulness of some global classification models is limited. Without the local expert knowledge some of the results can generate all kinds of incorrect conclusions (e.g. overestimated deforestation rates).

Something that we recently found out is the need to conduct workshops with the communities, professionals, citizens and government officials that live in the areas where a model is going to be applied. This is really important, it will highlight a lot of issues we might have overlooked.

I understand the importance authors are giving to CO2 emissions. we always try to improve the performance and the size of our models, but sometimes what we really need to ask ourselves is: Is this model really needed? Does it actually solve a problem?

You may have all the technology in the world but sometimes "on the ground" issues will make an AI solution completely useless. I would recommend conducting feasibility analysis before anything.