If satellite imagery has been freely available for at least the past 25 years, why isn’t everyone using it yet? Digital Harvest thinks they have the answer.
In 2014-15, Digital Harvest was a crop aerial imaging company. The company flew custom sensor payloads on aircraft to produce high resolution imagery. Farmers found the data “interesting”, in a few ways. If imagery was taken and analyzed at just the right time, and quickly enough, a farmer might have been able to do something about it. For example, detecting “skips” could allow limited replanting operations to be done to improve yields.
Most of the time however, regardless of the crop, imagery confirmed and quantified something the farmer already knew. Perhaps they have some low spots that create hydrology problems, maybe the edge exposed to wind has crop health issues, or there is a band of low organic soil through the field where yields are poor. Generally, farmers know what action could alleviate an issue; level the land, plant or build a windbreak, vary fertilizer application. These insights have quantitative value, but they are not revolutionary.
Furthermore, an indicator that looks problematic in one image may be a temporary condition. Perhaps a bad NDVI value in one part of the field isn’t a problem at all, that zone has a history of good yield and it’s just a few days behind on canopy closure. Making a good assessment requires viewing a time series of multiple images. Even after the data is gathered, it costs an organization significant time and resources to make use of it, and it still may be subjective to interpretation. Using imagery at scale comes with a high “interpretive burden”.
So what was missing? Digital Harvest found that almost all curiosity in more detailed analytical data had a common source: understanding, and predicting yield. Precisely quantifying the relationship between crop factors and yield outcomes is key. If there was a way to perfectly predict yield outcomes under any combination of inputs, an optimal solution that scales from individual plants to entire regions could be algorithmically determined and used to maximize profits.
Perfect prediction is practically impossible, but how accurate could it be? Would that be good enough? One thing was clear, any good predictive model had to be crop specific, there was no way to universally predict yield for all crops. Sugarcane, a robust C4 photosynthetic crop with a useful yield tightly related to it’s above ground biomass was a perfect crop to start with. Digital Harvest spent nearly 2 years performing projects for sugarcane growers on how yield prediction is done by farm managers, and how it might be done with the vast data sets available in the modern era without the need for the user to interpret it.
The results showed that a multi layered biophysical model and supervised machine learning approach allowed field level yield predictions with under 8% error (MAPE). When summed over many fields, seasonal errors of 1% for a region of mill were practical. These could be made months in advance with low cost satellite data, daily weather data, and minimal details about the crop. For many mills, yield prediction of this level of accuracy has significant financial planning value alone.
This error was low enough to start optimizing harvest schedules. Digital Harvest found that over a multi year term, in a benchmark case covering 80k hectares in Florida, USA, that an extra 2.5 tons of fresh cane per hectare (with the same average sucrose fraction), could be gained by just changing the order of harvests and planting to take growing time away from crops that approach their productivity potential and giving it to crops that would put on more biomass with that time. Such an optimal solution could be done without purchasing new equipment, or significantly increasing chemical treatments.
The next challenge is to package the analysis into a product that can be delivered to sugarcane growers all over the world, while maintaining the ability to customize the problem space to unique environments. The layered biophysical modelling and machine learning approach offers an opportunity to quickly start with “rough” biophysical yield predictions, while training data is gathered and processed to calibrate the AI layer to reach the 8% error level. Digital Harvest is hopeful that this approach will resonate with sugarcane growers and processors worldwide.
Digital Harvest is conducting a trial program to work with a limited number of sugarcane mills in each new country.
If you are interested in knowing more, share your details here.
Comments