Farmers across the United States use different systems of irrigation for different fields. Understanding their irrigation system is key to fully understanding their operational needs.
That's why we're excited to unveil our new irrigation data that will be available in all our applications this week. This data will provide deeper insights for you to better understand and serve your farmers.
For decades, people have used remote sensing to evaluate irrigated areas. We take that a step further to both identify and classify farmers' irrigation systems on a field-by-field basis. Learn more about how we conduct this analysis to provide you with better insights on the farmers you serve.
Overview of the process
We build our irrigation data using remote sensing and maps to place land into four irrigation categories:
- Pivot irrigation
- Non-pivot sprinkler
- Other irrigation (flood, micro, etc.)
For our purposes, irrigated land is defined as cultivated crop locations receiving application of water means to offset precipitation shortfalls during the growing season. This definition includes water sources like surface and groundwater deliveries -- as long as human intervention is involved in moving the water.
Determining irrigation criteria relies on a number of criteria. First and foremost is the availability of water resources and their interaction with vegetation. We also include climate, resource availability, crop patterns and technical expertise in our data collection.
The graphic below explains exactly how all this data fits together in the analysis:
The model relies heavily on monitoring changes in vegetation indices over the course of the growing season. We specifically analyze this data during periods of rain and dryness.
It's important that the timing of the image capture be precise. This helps to distinguish irrigated crops from each other, as well as from other land cover types.
This analysis does a good job at detecting vegetation activity over time through spectral (color) and temporal data. Spectral data examines the vegetative growth, while temporal data helps us determine the rate of growth.
Identifying irrigated vs. non-irrigated fields
Our irrigation data model is similar to our buyer persona model. In this case, we took a set of known data (referred to as a training data set) and ran it through a predictive model. Based on that training data, the model was able to determine and weigh key factor that indicate the irrigation status of a particular field.
Here are some of the findings of that process:
- In general, irrigated fields have higher vegetation index values than non-irrigated fields
- Most irrigated fields have a low year-to-year variation in vegetation index due to climate dampening
We used a tree-based machine learning model -- one widely used by the machine learning community -- to separate irrigated from non-irrigated land.
In addition to classifying land by irrigation type, the model returned a confidence interval for each classification. Of course, as with any predictive model, it’s important to test and refine the model to enhance its accuracy. In this case, the result is a cleaner and more accurate final irrigation map.
Finally, we evaluate the results and test them for accuracy in two ways:
- Use ground-truth observations to create a statistical estimate of the map's accuracy.
- Comparison of area estimates made from the irrigation map with those reported by the USDA at the county level.
After achieving confidence that the data is accurate, we then will deploy this data in our applications.