Research & Development

Absolute Sense™ from Leonardo

“Decision Support That Makes Absolute Sense™ ”

Normalised Difference Vegetation Index (NDVI) is a useful indicator of crop health but we believe we can provide deeper insights into your fields. By creating algorithms that work with remote sensing technology we can give you information inside the crop to help you make great decisions. Since 2014 we have been developing absolute mapping algorithms for primary macro nutrients, targeted initially at winter wheat.

Nitrogen, Potassium and Sulphur in-crop mapping algorithms are now available through Precision Hawk’s PrecisionMapper for the MicaSense RedEdge and Parrot Sequoia sensors.

In comparison to NDVI which only has an R2 of 0.26%
 
MicaSense RedEdge N uptake R2 is 0.94978%
Benefits of the Nitrogen In-Crop Map:
  • Instant view of nitrogen distribution in your field.
    • See in-field nitrogen variation.
    • See the distribution of the crop's nitrogen uptake.
  • Combine with other in-crop nutrient maps and/or soil maps to get a fill, layered picture of information.
    • Supports planning deicisions.
  • Values provided in kilogrammes of nitrogen per hectare.
    • Additional information to support agronomic decisions for calculating fertiliser applications.
Two colour profiles are supported (see figure 1):
  • Equal Area uses a dynamic legend, enabling users to see more detail across the field, highlighting where the variation in the field is at one time. This option is best for supporting fertiliser management decisions.
  • Equal Spacing visibly useful when comparing against other fields and also previous in-season maps of the same field across growth stages.
Figure 1 – Two Colour Nitrogen Maps
How Do We Do It?
The complex models created require two types of data. The first data from two high-end hyperspectral sensors that cover a wide range of the electromagnetic spectrum. Similarly: and the second is data representing empirical crop measurements. The hyper spectral sensors capture information from the visible through to short wave infrared, namely, 400nm through to 2500nm and attain a higher resolution of typically between 3nm and 6nm. Whilst these algorithms have been developed using hyperspectral data, equivalent multispectral sensor exploitation is achieved via a conversion utility we have created in support of the sensors widely used in agriculture today, namely the MicaSense RedEdge and Parrot Sequoia (with provision for other sensors planned to be released at a later date).

The rationale for using these high-end sensors during the research and development phase is to allow Leonardo to subsequently apply complex techniques to simulate any multispectral sensor available on the market today. Moreover, any loss in predictive accuracy when simulating a multispectral sensor has been found to be marginal. Taking this approach ensures a cost effective and timely route to market for the Absolute Sense™ products. Thereby, Leonardo is provisioning for any new sensor that is developed in the future, to be simulated by application of a conversion utility and historical hyper spectral data.

The following image shows the sampling plan for a trial field at a single growth stage – approximately one sampling point per hectare (see figure 2).
Figure 2 - Sampling Plan for Trial Field
The methodology for creating models follows essentially three phases. In the first, airborne data is captured, immediately followed by phase two, where a team of ground truthers are sent into the fields to gather crop samples. Each crop sample is individually tagged and then taken away for detailed lab analysis. In addition, the individual sample point location Global Positioning System (GPS) is accurately recorded and used for subsequent alignment with the airborne captured data. The final phase is the creation of the algorithms.

In order for models to be created to cover the full growing season, the data gathering process is repeated encompassing all key growing stages (both remote data gathering from hyper spectral sensors and ground truthing data) and for each growth stage a different sampling plan is adopted. The image below shows an airborne image after ground truthing has been carried out. The sample points are easily seen from this high quality aerial image. (See figure 3).
Figure 3 - Ground Sample Points

Field Selection
Each year, a number of trial fields are selected. These trial fields are managed by different land owners under guidance from agronomists, are guaranteed to have different soil types and winter wheat varieties. Through the volume and variability of sample data captured since 2014, the stability of the algorithms comprising the Absolute Sense™ product line is ensured.

Ground Truthing
A number of measurements are performed in the field and in the lab. These include: crop height, growth stage, fresh weight, dry weight etc. In addition, a representative sub sample is used for detailed nutrient analysis where key nutrients are measured, including: Nitrogen, Sulphur and Potassium.

Algorithm Creation
Post application of the conversion utility, transforming the hyper spectral data into representative multispectral equivalent data for the targeted sensor, the sample spectral profiles identified by the exact GPS positions of the ground truthing points are extracted. These are then aligned to the corresponding ground truthing results where Data Mining techniques are used to create the Absolute Sense™ mapping products. Leonardo has a highly specialist team of data analytic experts who carry out the complex analysis of multi-source data. The team has a wealth of experience in applying these advanced techniques to a wide variety of domains – including medical, military and agriculture.

MicaSense RedEdge Nutrient Models

Model creation uses multiple years of data covering key growth stages comprising both remotely captured and ground truthed lab analysis results.
The information below pertains specifically to those models created for use with data collected using the MicaSense RedEdge camera during Growth Stage 31, to calculate levels of:

  • Nitrogen (N)
  • Potassium (K)
  • Sulphur (S)

Below shows the model performance and application to unseen test data that was collected independently of the model generation. The graphs demonstrate the alignment of newly observed data to that of the corresponding model. This provides confidence that the model will accurately predict nutrients with other comparable data sets.

For each model, the R2 result is shown, along with the calculated predictive accuracy.

The MAPE (Mean Absolute Percentage Error) of a model will be referred to as a measure of the predictive accuracy.

MAPE (Mean Absolute Percentage Error) calculates the mean absolute percentage error (deviation) function for the forecast and the eventual outcomes. This is then used to express the accuracy of the outcome as a percentage of the error. (1 – MAPE) * 100 is used as a measure of the percentage of accuracy, rather than the percentage of error (see figure 4).

MAPE is calculated using the following equation:

Figure 4 - MAPE - Measure of Predicted Accuracy

The plots in Figure 5, Figure 6 and Figure 7 show the model performance against ground truthed data for Nitrogen, Potassium and Sulphur, measured in kilograms per hectare (kg/ha). The models have been created from trial data captured over several seasons of winter wheat growth at stage (GS) 31.

The plots in Figure 8 and Figure 9 show a crop’s Nitrogen uptake using NDVI measured in kilograms per hector(kg/ha). The plots have been created from trial data captured over several seasons of winter wheat growth at stage (GS) 31.

Nitrogen
The R2 of the model is 0.9498, with a predictive accuracy [(1 – MAPE) * 100] of 78.914%

Figure 5 – Calculated Nitrogen (Predicted) against Nitrogen Uptake (Ground Truthed)

Potassium
The R2 of the model is 0.8649, with a predictive accuracy [(1 – MAPE) * 100] of 70.229%

Figure 6 – Calculated Potassium (Predicted) against Potassium Uptake (Ground Truthed)

Sulphur
The R2 of the model is 0.7987, with a predictive accuracy [(1 – MAPE) * 100] of 76.823%

Figure 7 – Calculated Sulphur (Predicted) against Sulphur Uptake (Ground Truthed)
NDVI against Crop Nitrogen at GS31

Example 2016 Season GS31 Data

Figure 8

Example 2017 Season GS31 Data

Figure 9
Top