As the word "AI" sweeps across the globe, it's time to deep dive and understand how we use the power of artificial intelligence and machine learning tools to leverage our in-field IoT data streams, and help growers make smart irrigation decisions in real-time.
Avi Braun, who heads the development of our AI-based Algorithm explains in this post.
No Two Plots Are The Same
"At Phytech, we measure and analyse tens of thousands of agricultural plots every day. This provides us with a wealth of data to learn from regarding the expected behavior and performance of various plants in different climates around the globe. However, we also recognize that farming plots can exhibit different behaviors, even when all known variables are equal. In other words, two plots with the same crop type, age, and nominal soil type may still perform differently due to hidden factors that are not apparent to farmers or us.
When it comes to recommending irrigation timing (and soon also quantity!), we learn from both the collective behavior of similar plots and the performance of each individual plot.
Combining Data From The Plant, The Soil and The Climate
For our Dynamic Soil Thresholds, our approach involves utilizing the data generated by our soil and plant sensors, as well as climatic data. We assess the plant's performance at the end of each day and correlate it with the soil moisture level for that day, taking into consideration the weather conditions and the water demand by the atmosphere on that particular date.
The plant dynamic refiil-Line continuously adapts itself:
Training The Machine
By analyzing the entire system—soil, plant, and atmosphere—we can gain insights into the soil moisture conditions that are more likely to cause stress to the plant. To achieve this, we apply our customized online training algorithm to continually retain and refine our estimation of the threshold as the season progresses and the plant undergoes different phenological stages.
For the Soil-Prediction algorithm, we apply Supervised Machine Learning techniques where we learn from the ‘history’ of each soil sensor what would be the soil-moisture level in the future if it will not be watered.
Learning From The Past To Predict The Future
For our predictions to be accurate we first validate and clean the data of each sensor and correcting its reading to account for the influence of temperature on the measured values. We then use Ensemble Regression Trees model to predict the rate at which the soil will dry in the coming days. In our model-training and predictions we also take into account historic and future weather conditions. Similar to the Dynamic Soil Thresholds, we use on-line training to keep refining and improving our models every day as collect more data.
1 Day To Refill Prediction
