Walk through a tomato field in midsummer and no two plants will look exactly alike.
Some are just beginning to flower. Others are heavy with ripening fruit. A few may already be showing the first signs of stress from heat, drought or disease.
For farmers, recognising these changes early can make the difference between protecting a harvest and losing part of it. Yet monitoring hundreds or thousands of plants throughout an entire growing season remains one of agriculture’s most time-consuming tasks.
As climate change continues to make growing conditions more unpredictable, that challenge is only becoming greater. Tomato production is expected to decline in the coming decades, making efficient crop monitoring increasingly important for maintaining yields.
A new study by AgRibot partners from the National Research Council of Italy (CNR) and ALSIA explores a different way forward: AI that doesn’t simply learn once, but continues learning as crops and growing conditions change.
Moving beyond static AI
Today’s AI models are typically trained using a fixed dataset. When new crop varieties, growing stages or field conditions emerge, developers often need to collect new images and retrain the model from scratch—a process that requires considerable time, computing power and manual data annotation.
Instead, the researchers developed an incremental learning approach.
Rather than rebuilding the model every time new information becomes available, the system gradually updates its knowledge using smaller batches of data, allowing it to adapt while retaining what it has already learned.
Smarter learning, better performance
The team evaluated several versions of the widely used YOLO object detection model before identifying YOLOv11m as the best balance between accuracy and computational efficiency for tomato phenotyping. They then applied their incremental learning framework to improve the model over successive training stages.
The result was a model that achieved higher detection accuracy, more stable predictions and slightly faster processing than conventional training methods. It also successfully adapted from controlled laboratory conditions to real tomato fields using only a relatively small amount of additional field data.
Interestingly, the study also found that adding increasingly complex AI components did not necessarily improve performance. In this case, the simpler model proved to be both more accurate and more practical for real-world agricultural applications.
Why it matters
As agricultural robots become more common, their ability to understand what they see is just as important as their ability to move through a field.
AI systems that can continuously adapt to new crops, environments and seasons could reduce the time and cost needed to deploy precision agriculture technologies while requiring far less manual data collection. That means farmers could benefit from more reliable crop monitoring tools that improve over time rather than becoming outdated after a single growing season.
While further research is needed before these systems are ready for widespread deployment, the study demonstrates an important shift in agricultural AI: instead of teaching machines everything at once, researchers are finding ways for them to keep learning alongside the crops they monitor.