For a company dedicated to the cultivation and distribution of packaged and prepared lettuce, harvesting becomes a decisive moment in its business. When it comes to buying fruit and vegetables, appearance plays a fundamental role in the purchasing decision. Whether the packaged product is in the best condition depends to a large extent on the harvest.

Deciding when is the optimum time to harvest lettuce was so vital for Florette, the leading company in its sector in Spain and Portugal, that a large part of its investment went into this task. 

Lettuces are harvested between 60 and 80 days after planting, for instance, a period of 20 days in which human inspections had to be carried out daily until the right time for harvesting. This matter is translated into thousands of working hours per year.

Optimizing this monitoring was vital for Florette, but the IoT sensors installed on its crops did not provide enough data to make a decision. That's when Artificial Intelligence came into play. Specifically, one of the solutions offered by Google: BigQueryML.

Predicting when to harvest

Knowledge and tradition have always played an important role in agriculture, where years of experience led harvesters to make the best harvesting decision. With the advent of Artificial Intelligence, it is possible to improve this process: scientific data is obtained from reliable sources, over a period of time, and processed to obtain patterns and trends. 

This was the work that was carried out with Florette: historical sowing and harvesting data for crops were taken and matched with weather records and forecasts.

These sets were housed in a BigQuery repository where, thanks to its Machine Learning functionality, ML models were created. By not having to move data between tools, the speed of development and efficiency of the solution is increased. 

"Thanks to ML models, the number of human inspections has been reduced by 50%".

In order to obtain predictions, two models were programmed:

• Optimal harvesting of crops based on meteorological data.
• Prediction of kilograms harvested per crop

Once built, these two models are continuously fed with new data and periodically retrained to optimize their learning, giving increasingly accurate predictions.

Smarter harvesting

The implementation of prediction models has enabled Florette to reduce the number of human inspections by half, resulting in significant cost savings and a simplification of the entire production process.

As far as the accuracy of the models is concerned, the prediction accuracy exceeds 85% of the cases. Such has been its success that this solution, designed for iceberg lettuce cultivation, has been extended to the company's other crops.