The industry has made significant progress in using satellite imagery to monitor fields, especially regarding the amount of data available. While we have solved the problem of getting enough data, we have inadvertently created a new problem – “all the data in the world doesn’t matter if it’s not easy to disseminate to make better decisions.” And this is where change detection comes into play.
In 2019, Geosys delivered an average of 24 cloud-free maps per field in the United States (with more than half of those maps delivered less than 24 hours after acquisition). Since the growing season lasts about 120 days from planting to harvest, we basically delivered a map every 5 days per field. When you consider an agronomist needs to monitor as many as 1,200 fields each season, that could potentially be 28,800 maps – that is a lot of data to consume!
Geosys has several data analytics to help users make better decisions when it comes to focusing their attention on the right fields. Field benchmarking – which compares vegetation index measurement (NDVI or EVI) of one field to that of similar fields to quickly see which fields are performing higher or lower than average, and how those fields are trending – helps users quickly understand which fields are building a higher yield potential crop than average. The next stage is to automatically identify when change takes place within a field. Therefore, we developed analytics to help users better see and understand the level of change, aka Change Detection.
In 2018, we conducted a series of pilots to test our analytics for change detection. Users were very pleased with the results, so we are launching Change Detection 1.0. This initial launch helps users quickly determine which maps to look at for a given field by providing a score from 0-10 indicating the amount of change for a given field. So, instead of flipping through a library of maps to understand what is going on in each field, Change Detection 1.0 tells you “look at these two maps.”
How Change Detection Works
Change Detection 1.0 is based on three components of change using the fields NDVI values:
Trend: Did the field experience the same change? Did all pixels increase by the same amount (small change) or some pixels increased and some decreased (large change)?
Type: Is the field variability the same? Are the good spots of the field still good and the bad spots still bad (small change) or are there new good or bad spots which have emerged in the recent image (large change)?
Intensity: Is the overall NDVI distribution the same? Does the histogram of the NDVI values for the new image overlap with the previous image (small change) or is there very minimal overlap between the histograms (large change)?
The quality of satellite data is paramount to delivering proper change detection. Using “noisy” satellite pictures only creates additional clutter with false positives results. This is why we elect to work with trusted satellite data sources and process the data to deliver consistent measurements, through intercalibration and atmospheric corrections.
The following are extreme examples of the various types of change to help visualize the three components of change:
Trend We can see that part of the field is improving while other parts are getting worse. While Type and Intensity might have caught some level of change, correlation is needed to understand the extent of this change. The change index here would be 10.
Type Here we see a homogeneous field that has changed into one with a lot of variation. The change index here would be 10.
Intensity While we can visually see the change in these two maps, the change in Trend and Type are low because the entire field is getting better and variation pattern is staying about the same. This is expected. The change index here would only be a 6.
Best Uses for Change Detection 1.0
- Quickly see various rates of change and helps identify which maps to look at for a given field.
- Review a side-by-side comparison of the latest two images within the 5- to 15-day threshold and see how each of the components are affecting the change detection score.
- After evaluation the maps and charts, determine if further action is needed. For example, do you need to scout?
Why 1.0? The analytics have been tested in real situation (25,000 fields) in 2019 and we continue to learn, adapt and expand this offering. We are currently exploring how to decipher good change from bad change and how to leverage ground truthing data to qualify change with probable cause.