Artificial Intelligence in Agriculture
There are many possible applications of artificial intelligence in agriculture. Quite a bit has been written about the topic and much of it focuses on precision agriculture applications.
Artificial intelligence in agriculture is being applied to monitor crops, plant pests, identify damaging conditions via computer vision to detect disease, and many other applications. An application that has emerged recently includes predictive analytics and enabling agricultural robots to perform specific tasks.
The CropProphet objective is to create the most accurate national yield and production forecasts possible for corn, soybeans, and winter wheat. This is achieved by quantifying the impact of weather on crop yields.
CropProphet supports agricultural futures trading and risk management. We do not create precision agriculture applications but we do offer risk-free trials of our product.
This post discusses an experimental application of artificial intelligence in agriculture as applied to the CropProphet corn yield forecast product. CropProphet is a machine learning-based yield/production forecasting system that finds the optimal relationship between many years of weather data and end of season USDA NASS crop yield and crop production reports.
Let’s highlight two definitions: machine learning and artificial intelligence (AI).
We interpret machine learning to be the application of computer based computations to scale a particular calculation. Until recently, machine learning has largely comprised the application of different linear methods to conduct an analysis of some kind.
CropProphet is a machine learning-based system because we forecast yield and production for thousands of US counties by creating county-level prediction models. Machine learning enables us to do this at scale.
Artificial intelligence (AI), also called deep learning, is the application of machine learning techniques enabling non-linear relationships to be developed. In recent years, supported by the ever-increasing scale of the availability of computing resources, the number of techniques associated with AI has exploded.
Artificial Intelligence and CropProphet
In this post, we are discussing an analysis of corn yield forecast and corn production forecast risk analysis using artificial intelligence. It discusses the application of a machine learning approach (k-means clustering) and an equivalent artificial intelligence approach (self-organizing maps).
In a recent post about how we create our crop weather forecast product, we highlighted that CropProphet provides a yield forecast based that quantifies the impact of the weather forecast models. For example, the ECMWF model weather forecast is actually 51 different possible realizations of future weather conditions.
Everyone knows weather forecasts have uncertainty. The uncertainty of those forecasts is measured by producing 51 different weather forecasts. It’s called an ensemble of forecasts. The model is exactly the same each time. The only thing that’s different is the initial conditions of the model (i.e. the temperature and winds, etc at the time the model is started).
CropProphet has quantified the impact of weather on crops. Because of this, we can apply our model to the weather forecast data to create a yield or production forecast. Using the ECMWF model every day, we update the yield forecast model. When this is done we are actually creating 51 different yield forecasts (and 21 for the GEFS, and 16 for the CFS) from the ensemble weather forecast data.
What does 51 different yield forecasts look like? See below.
Crop Forecast Risk Analysis
In our recent post regarding crop weather forecast impacts on yield, we highlight this methodology as a means to prepare a “best case” and “worst-case” analysis based on the crop weather forecasts. However, the best and worst-case forecasts are literally the extreme values of the corn yield forecast in the 51 member ensemble.
K-means cluster analysis in agriculture
Given 51 different forecasts, each a different possible realization of future corn yield, it would be helpful to know if some are more similar than others. It might help put all of this data in context.
Enter the k-means clustering process. The clustering process finds each of the corn yield forecasts that are more similar to each other when compared to all the other forecasts. An example of the clustering of CropProphet corn production forecasts from June 15, 2020 is shown below.
Note that it shows what the 15-day change in the production could be and the probability of that forecast outcome verifying as correct.
It helps to analyze the risk in the corn yield forecast. This method, however, is a linear based analysis. It’s certainly valuable but it might be possible to provide a deeper analysis.
Self Organizing Maps in Agriculture
The same analysis can be conducted with a self-organizing (SOM) map analysis. The possible advantage of a SOM is that it uses a non-linear neural net-based process to find clusters and may create a better analysis of the risk of the forecasts. We are experimenting with this methodology today to try to better understand the probability of outcomes of weather forecast impacted yield and production forecasts.
The SOM analysis above is based on the same 51-member ECMWF corn yield forecast change analysis presented to the k-means cluster. The same data for risk analysis can be extracted from the process.
This is our first attempt to apply artificial intelligence to CropProphet. We are generating these analyses each day CropProphet is updated. We expect the SOM process will better capture extreme situations because of the non-linear nature of the analysis. However, the product is experimental is we are not distributing the data to customers.
We continue to try new methods to improve the quantification of weather risk to corn, soybean, and winter wheat crops.
Try CropProphet – we offer risk-free trials.