Commodities: Challenges of Predicting Crop Yields
Predicting the yield of crops for a farm, country, or even a region is both important and challenging – in terms of the data required and the levels of analytics involved. Historically many crop predictions have been little more than 30% accurate. Fortunately, the accuracy in predictions of crops, including sugar and rice, can now exceed 95%.
This first article will outline the challenges of crop predictions, the second article will explain the solutions that enable high levels of accuracy in predicting crop yields.
Prediction of Crops – Challenges
Crop Predictions are important for governments, farmers, and traders – a more accurate prediction of future supply allows for better policies, better profits, and better pricing.
The concept of using satellites to predict crop yield is very simple: Take a picture of a field, look at the crops growing, measure the area of the field and predict the yield. However, there are substantial challenges to this, including:
- Scale – global
- Size – small farm to large agri-business
- Colour – it’s all green
- Phenology – Identification of relevant crops
- Area v Yield – Giving the area does not give the yield.
The first and most obvious challenge with taking pictures by satellites are clouds. Farmland is, quite deliberately, not in hot dry areas. Regular rain, and associated clouds, means that simply getting pictures of crops can be challenging. In England, for example, it would be almost impossible to find a day without cloud cover over some of the farmland.
For predictions to be effective this has to be conducted on a national or international scale – i.e predicting the yield of a single farm isn’t sufficient. The predictions have to over a wide area to be of any value.
The identification of crops, using satellites orbiting at 800km, is based on a variety of metrics and observations. The most common one used is to look for “green” fields. The challenge with this method is that many areas are green. Grass, forests and other crops – all appear green and look very similar to the required crops. For this reason, the challenge is to filter out the irrelevant “green” and be left with the relevant – the actual crops.
If all the farmland was a single large continuous farm the analysis of the crops would be far easier – however, farms tend to be a mixture of small farms and large agribusiness varying in size by the owner, the crop and the economics of the country. Identifying different farms, and what they are growing at different scales becomes complex and prone to error – i.e. Farm A, Size B, is growing Crop C in Fields D, E and F and Farm X, is growing Crop Y in Field Z. These different fields and different crops all need to be identified and allowed for in the overall calculation.
To understand a particular crop, e.g. sugar, it is not enough to just differentiate the green fields of farmland from the green fields of woods and grass. There has to be correct identification of the relevant crop. This is complicated as some farmers grow multiple crops – i.e there may be crops of sugar and cassava, which look similar and can look identical to a satellite from 800km away. Knowing that a particular farm has 100 hectares of crops is not useful, you must know if he has 80 hectares of sugar and 20 hectares of cassava.
Area v Yield
Finally, once the crop has been identified and the area calculated, this only gives the area and not the actual yield. The yields will vary depending on a variety of factors including weather conditions and the health of the plant.
Predicting the areas of crops being grown across a country, under cloudy skies, with variations in farm practice, in a wide range of different size fields is, at best, complex. Calculating the predicted yield, from the predicted area of crops – i.e. going from the number of predicted hectares of sugar to expected tonnes – adds additional layers of complexity.
The next article looks at solutions for these challenges – which can drive high levels of accuracy.