How To Quantify and Predict Input Performance in Crop Research
The information needs of agricultural input suppliers require a thorough understanding of how inputs, environment and management practices all combine to produce yield.
This means being able to isolate variables such as specific inputs or crop management decisions, and find the statistical link to crop performance or yield amidst a myriad of factors.
SlantRange joined DroneDeploy and Beck’s Hybrids for a recent webinar in which CEO Mike Ritter discussed how our technology enables agricultural input suppliers to understand how variables impact crop performance.
This video excerpt provides an overview of how our technology helps researchers quantify and predict input performance. Use cases and examples cover analyzing stand and vigor, quantifying response to inputs, and measuring growth stages.
The transcript, lightly edited for clarity, follows with visuals from the presentation. The full webinar, featuring presentations from Jim Love of Beck's Hybrids and Anna Schneider of DroneDeploy, is available from DroneDeploy.
My name is Mike Ritter. I'm the CEO of SlantRange, and I'm happy to be here with DroneDeploy and Beck's Hybrids today.
We are a remote sensing and analytics company, but we work in a slightly different area where the needs are a bit different. What I want to do today is introduce some applications, some cases in technology, which are focused more on agricultural inputs supply as opposed to agricultural production.
The hope here is that some of you listening in, may see a fit with your needs. The technology looks the same on the surface, but it's really quite distinct in the application. So that's what we're going to discuss today.
Listening to Jim's presentation, we're hearing a lot of keywords about early detection of emergent problems, responsive in-season management decisions, optimizing individual field performance, and those are regular challenges for ag producers.
The needs of the ag input supplier are a bit different. What we're trying to work on here is understanding how inputs, environment and management practices all combine to produce yield.
That really boils down to a few things. Statistically attributing crop performance or yield to specific inputs or management decisions amidst the myriad of variables that are affecting productivity, and amidst soil weather, inputs, selection, and timing, and rates and all those things, one of the real challenges is figuring out how the adjustment of one input can really affect the outcome. That's a difficult problem.
Once you can do that, building a base of data that demonstrates that capability, so that you can use that information to help build models for forecasting and input prescription.
The technologies are similar, but they differ in your accuracy, sensitivity types of measurements and the costs involved. I'm going to go through some of that briefly here today.
A little bit of background on our company: We founded in 2013. We are located in San Diego, Southern California, and we first began servicing input suppliers in 2017.
At that time, we just ran one trial location and analyzed about 20,000 plots that year. As of today, we're approaching about 700 trial locations with well in excess of 3 million plots analyzed.
Over that period of time, we've seen a really rapid adoption of the technology amongst input suppliers. The question is why.
First, it's automation of labor and labor-intensive tasks. What input suppliers are learning is that the improved statistical confidence that you can gain in how individual inputs are performing in development trials reduces the scope of trials that need to be conducted in order to bring in a new input to market.
Ultimately, what that does is it combines to reduce the time and cost of bringing inputs to market.
Tools for Quantifying & Predicting Crop Input Performance
I'll give you a little bit of a background on the technology here. Our goal is to achieve less than 1% error in the delivered metrics across a number of variables that we measure throughout the growing season across a number of different crops.
Population statistics, or emergence, are very important for different reasons than what Jim talked about in his presentation: plant morphology or leaf shape that contributes to leaf canopy structure, plant pigment concentrations, growth stage metrics, like maturity and flowering.
We work heavily in small plot trials. When you work in that type of environment, the spatial accuracy of your measurements is critical. You're also analyzing data that's coming in from hundreds of locations potentially distributed globally.
Aggregating all that information to be analyzed together over time and space becomes a bit of a complex bookkeeping challenge. The technology that we deliver really spans remote sensing, in-field edge computing, as well as the cloud computing piece.
On the sensing side, we've developed a spectral imaging device that has what we consider super critical spatial resolution, which means we can isolate segments and measure individual leaves within the field.
It's done with very narrowband spectral imaging and uses an onboard solar calibration component to normalize measurements so we can make comparisons over time. It's very sensitive and accurate to subtle changes in plant condition. It also incorporates LIDAR and RTK for precision navigation. Again, it's critical for a lot of our applications to place those measurements to within 10 centimeters of truth on the ground.
There’s an edge analytics component, which is primarily created for rapid in-field quality assessments of the data and some basic scouting functions.
A lot of the heavy lifting occurs in the cloud analytics platform where we're aggregating, organizing data, and applying a full suite of analytics, modeling and forecasting.
I want to run through just a handful of cases here relatively briefly.
Use Case 1: Analyzing Stand and Vigor
Jim already was talking about the importance of stand measurements early-season — also important for the agricultural inputs for a little bit different purpose here.
In production, what you want is a quick assessment to make an input decision. For example, do I need to make a replant decision?
What we want to understand on the input supply side is to quantify seed or seedling vigor and assess performance across different input varieties and treatments. Another really critical aspect here is achieving separation of variables, so that at the end of the season, when you're looking at yield, you're able to separate what contributed to yield, how much of it was due to stand in early-season versus what might have been treated or happened later-season.
There's multiple steps involved here. First step, of course, is classifying, segmenting, and measuring individual plants in the imagery. Then we can produce a number of metrics like plant count; plant size statistics; emergence fraction; vegetation fraction, which deals with canopy structure; and weed metrics.
This is basically what the user interface looks like on a small plot trial of corn at about V2 stage. What we see here is, the central area of is a satellite background image with a heatmap overlaid.
Here, the variable that we're looking at happens to be plant size in square centimeters. Green are larger plants, red are smaller plants. And then gray areas, black areas are little to no stand at all.
Each pixel color represents the median plant size within that five-meter-by-two-meter plot. At the bottom left, there's a histogram of the plant sizes across all the plots. At the bottom right is an image chip of a single plot with green overlays for the corn plants, and then red overlays for weeds, which are a little bit difficult to see, but they're at the top end of that plot.
At the top of the frame, there's some buttons to select different data variables that were collected at the same time: emergence fraction, plant count, weed count. This is a user interface that is helpful for visualization.
All this data is available for download or API interface to pull into the input supplier's own analytics or their own platform.
So this approach is a little bit different from the standard counts that we see in production agriculture. When we're aiming for less than 1% error, what that means is in a plot that might have 70 plants in it, we can exceed our error budget if we miscount a single plant.
So it's absolutely critical we identify and properly classify individual plants for measurement. The spatial accuracy is required as well. If you misplace that aerial measurement by 10 or 20 centimeters, you might start overlapping another plot, and then once you do that the results become unusable.
All those features combined allow you to really quantify stand across large-scale trials like this.
Use Case 2: Quantifying Crop Response to Inputs
For input suppliers, it's critically important to be able to quantify the efficacy of their product to demonstrate overall grower returns and be able to target inputs for specific localized environments.
The challenge here, and I'm talking largely about crop protection, suppliers, crop nutrition suppliers, biologicals, the challenge for those companies is often demonstrating that their input performed better than another input given the myriad of variables always at work in any field.
So the challenge here is really to try to isolate and quantify the performance of an input subject to whatever treatment may be applied.
What the solution involved here is really looking at the time-based response of plant development in response to those inputs.
In the next slide, what we see is a similar layout. This happens to be a crop protection trial on corn. This is a bit later in the vegetative stages in corn.
It's a managed stress environment where stress is introduced, and then a treatment is applied to see how the plants respond to that stress. The heat map at center is what we consider plant stress, a four-band NDVI. It's a bit more sensitive to plant stress than NDVI. Green is relatively healthy, red is more stressed plants. At the bottom left, again, you have a histogram of all the plots across the trial. At right, you have a high-resolution image of the location.
What's different in this example is the line plot you see at the center at bottom, which is a time series of two different variables. I think this is vegetation fraction or canopy closure, along with vegetation stress. What we're showing is how an individual plot is responding to treatment over time.
So what's really different here is the ability for the input supplier to quantify that, and provide evidence of the efficacy of their input, as the crop is responding in real-time. This is all enabled by sunlight calibrated narrowband spectral imaging techniques.
Use Case 3: Measuring Crop Growth Stages
I'm going to give a couple of examples here briefly on growth stage measurements, which are important for a couple of reasons. Some inputs are designed for specific localized growing environments, and timing of inputs is critical relative to the growth stage in some crops as well.
What we've done is developed some time-based metrics that are looking at how crops are developing to provide measures on the plant growth stage.
Canola Flowering Dynamics
The first example is a canola trial. Again, small plots. In this particular case, we're using the spectral signature of canola flowers to measure the flowering fraction development in each plot.
In this particular heat map at the center, which is percent of flower, green shows a higher percentage of flowering and red is less. The time series at bottom shows the flowering development for the selected plot that's shown at top right.
The ability to track flowering development in canola is particularly important for the timing of inputs. It’s also used as a metric for yield forecasting.
One of the takeaways here is the ability to isolate specific variables for measurement is one of the critical aspects of using aerial measurement techniques like this.
Soybean Maturity Ratings
The next example in the next slide is a soybean trial. Soybeans are bred for different growing environments based on the length of the season. Historically obtaining soybean maturity ratings, as they're called, is an extremely labor intensive and subjective process.
Ultimately, what we're trying to do here is breed seeds that are going to optimize yield based on the length of season.
Again, similar layout and the data you're looking at here. The time series at bottom is the measure of maturity for the selected plot. What we're showing is that in the pop up, the data allows us to actually measure the days to maturity, which in this case was 121, and the date that that maturity was reached.
Just in closing in our section, this has been a very brief introduction, and then just cut across a few samples. What we're doing is working with input suppliers to help improve their understanding of how different inputs perform within relatively complex growing environments to understand how they contribute to yield individually.
Once you can do that, you can begin optimizing those different inputs and management practices. The data that is created in the process provides input supplier tools to help market those inputs to their customers.
Ultimately for the grower, it results in higher performing inputs and practices. So one question we always get is do these tools apply to ag production? They absolutely do. That's a deeper conversation.
If you're interested in learning more, we're happy to talk. Just get in touch with us. Thanks to DroneDeploy and Beck’s Hybrids for participating with us here today.