For protected cropping growers, yield prediction is an important tool for both short-term response and long-term planning. While many growers aim for high prediction accuracy, that accuracy is not always consistent.
Drawing on insights from our Beyond Yield Prediction: The Importance of Climate and Context in Tomato Production whitepaper, this webinar makes the case for a more holistic approach to yield forecasting. It shows how climate, plant, and contextual data can help growers interpret predictions more effectively and make more confident decisions.
Presented by
Welcome to today’s webinar. We’re going to cover three things. First, we’ll look at the singular focus on achieving high accuracy in yield prediction and ask whether that is the right approach. While yield prediction is a valuable tool, and we strive for high accuracy, it is worth questioning whether accuracy alone should be the main goal.
Then, we’ll discuss yield swings, what they are, and what causes them. Finally, we’ll talk about interpreting yield prediction with greater context to improve consistency and confidence in outcomes.
This research, produced by Dr Tharindu Weeraratne and Dr Mpatisi Moyo, is also available in the Beyond Yield Prediction whitepaper.
We’re covering yield prediction and forecasting today, but let’s start with a quick refresher on why we do yield prediction and what we aim to achieve with it. Yield prediction drives both long- and short-term decision-making. In the long term, it helps ensure enough is produced over the season to meet contracts and revenue targets. In the short term, it helps steer plants to meet immediate contract requirements. Yield prediction allows growers to understand when they may not meet targets and enables them to steer crops towards desired outcomes. If production is likely to be too high or too low, growers can adjust accordingly. Yield prediction increases confidence in planning and fulfilling contracts.
But how does yield prediction work? At WayBeyond, we use a machine learning model that uses historical data to find patterns and predict future outcomes. The model looks at patterns in yield and environmental conditions and applies them to current greenhouse conditions to forecast future yields. This approach is valuable, but it does have challenges. It requires a large volume of historical data, which is not always available and may be siloed. Changes in climate and weather patterns can also make historical data less relevant.
One significant issue is the variable accuracy of yield predictions. Sometimes predictions are highly accurate; at other times, they are not. Accuracy can range from 65% to 100%, and it changes week to week. This variability is a challenge across all methods of yield prediction.
So, while growers aim for high accuracy in yield prediction, the question remains: is focusing solely on accuracy the right approach? This presentation argues for a more holistic approach to planning and forecasting, where yield prediction is one of several tools. It is important, but it should sit alongside a better understanding of the environment, plant state, and management practices.
To understand the impact of variable accuracy, we looked at 20 cycles of tomato harvest data. These cycles were grouped by average yield prediction accuracy, and then the number of yield swing weeks in each group was analysed. Cycles with the lowest average yield prediction accuracy had the most yield swings. This suggests that as the number of significant deviations from expected yield increases, the accuracy and usefulness of yield prediction decreases.
For both low- and high-tech growers, several factors can influence these yield swings. Understanding these factors is critical. Environmental data from before low and high yield swing weeks showed common factors including low outside night temperature, total light, and internal day-night temperature differences.
We also looked at plant data and found that plant state before a yield swing week is an important indicator. Before a low swing week, plants were often in a vegetative state. Before a high swing week, they were more generative.
So, what does this mean for tomato growers? Yield prediction is more effective when supported by relevant local insights. It is a valuable tool, but it should be used alongside an understanding of the growing environment and plant conditions. Rather than focusing only on maximising prediction accuracy, it is more useful to understand the broader context of greenhouse conditions. Investigating yield swings and trends can help growers anticipate and manage future production. Plant state is also an important indicator of potential swings, so monitoring plant measurements remains essential.
In conclusion, yield prediction is an important tool, but it should form part of a more holistic approach to planning and forecasting in protected cropping.