When AI encounters natural history: Morphological OTUs reshape our understanding of Earth's life

To understand the living world, scientists must first agree on what they are looking at. For centuries, this has meant sorting every leaf, beetle, and flower into a named species, a task that relies on experts who can recognize subtle differences in shape and structure. This system works well when specimens are rare and well-studied, but it breaks down when facing the sheer volume of life found in a single field survey, where most individuals cannot be identified to a specific name. To cope with this, researchers have often turned to molecular methods, grouping organisms by their genetic codes into units that stand in for species. However, these genetic approaches require collecting tissue samples and running complex lab tests, which is not always possible. Meanwhile, simple visual identification has struggled because it usually depends on already knowing the species names to train the computer. The challenge remains: how can we organize and count the diversity of life using only the pictures we take, without needing a pre-existing list of names or a lab full of equipment?

A new study offers a way forward by teaching computers to see the world the way a naturalist does, but with a much broader scope. The researchers developed a system called morphOTU, which creates groups of organisms based entirely on their physical appearance. Instead of asking the computer to memorize a list of known species, the system learns to recognize patterns in the shapes, textures, and structures of the organisms themselves. It organizes thousands of images into a continuous space where similar-looking things cluster together, forming what the authors call operational units. These units act as a practical layer for counting biodiversity, allowing scientists to measure the variety of life in an area even when they do not know the formal names of the creatures they are studying.

The team tested this approach on five different sets of data, ranging from images of flowers and wood anatomy to the bodies of beetles. In each case, the system successfully grouped the images in a way that matched the structure of real species. When the researchers compared the diversity counts generated by the computer against those made by human experts, the numbers were remarkably close. The system worked even when it was trained without any labeled species names, and it remained accurate even when the researchers provided very few examples of specific animals to learn from. This suggests that the visual patterns the computer found were not just random noise, but genuine reflections of biological reality.

To see how this would hold up in a real-world scenario, the researchers applied the method to a massive field survey containing 4,717 insects representing 269 different species across 12 distinct orders. In this complex mix, the system was fine-tuned using images from only 28 common species. Despite this limited training, the resulting diversity estimates were nearly identical to those produced by expert human labels. The computer calculated a diversity score of 3.75, while the experts calculated 3.53, a difference so small it indicates the machine had effectively captured the true variety of the community. The system did not just produce a number; it also showed the researchers exactly where the differences lay. It could point to specific features, such as the symmetry of a flower, the texture of a beetle's shell, or the arrangement of vessels in a piece of wood, proving that the groups it formed were based on biologically meaningful traits.

This work demonstrates that we can build a reliable framework for understanding biodiversity directly from the physical forms of organisms, without waiting for formal names to be assigned. By organizing life into these visual groups, scientists gain a powerful tool to quantify the richness of ecosystems in the field, alongside traditional methods, or in situations where species names are simply unavailable. The findings show that the visual world holds enough information to sort life into coherent units, offering a new way to see and count the natural world before, alongside, or in the absence of a formal species name.