Orchard and vineyard and systems are spatially and temporally variable but growers have to date only had complex, expensive and tedious strategies to measure the variability of the plants. As a result, the information growers have about the state of their crops is sparse and inaccurate, meaning they are limited in how they could manage their operations.
We present and discuss a strategy to measure the crop yield efficiently using imagery captured from vehicles driving through the fields. Our method is non-destructive and can produce dense measurements during the season by detecting within imagery using algorithms that recognize the shape, shading and color of the fruit. The measurements can be translated into georeferenced yield maps that capture the variability in the fields, and we have seen large variation patterns that raise a number of topics, such as the origin of the yield patterns, whether it be underlying soil variation or management practices, and also raises the question of how growers can leverage this type of information within their operations.
We study the issue of scaling the system to large vineyards and orchards through the use of optimal calibration and sub-sampling strategies. We have also studied how to efficiently reconfigure the system for fruit with different visual appearance. Our system has been demonstrated over several seasons in a variety of different production systems and validated against the true harvest measurements.
Stephen Nuske is a Senior Project Scientist at the Field Robotics Center at Carnegie Mellon University. His expertise is visual scene / object segmentation, visual tracking and visual localization for outdoor robotic systems including UAVs and ground-based agricultural vehicles.