Commodities: Crop Predictions Challenges

The challenges of predicting crop yields are significant, as outlined in the previous article. Traditionally these challenges result in low accuracy rates –  below 50%, for certain crops.  This article will look at how the accuracy rate can be increased to 95%+

Prediction Crops Challenges

The primary challenges for predicting crop yields are:

  • Clouds – limiting photography by satellites
  • Scale – The global scale puts demands on capturing and processing the data
  • Farm Size – Varying farm sizes –  from individual farms to giant corporate entities increases the demand and complexity of analytics. From small farm to large agri-businesses
  • Differentiating the green farmland from the green countryside (woods, grass, etc)
  • Phenology – Identification of relevant crops, e.g. separating out sugar from cassava
  • Area v Yield – Predicting the yield, based on the area of the planted crop.

These challenges are dealt with below

Clouds

Clouds limit what pictures can be taken, i.e. if a satellite flies over once a week and the farm is cloudy four days a week it’s more than likely to be cloudy when the satellite passes over.  This is addressed in three ways:

  • High revisit times
    • The higher the frequency of visits by the satellite the higher the probability of getting a good image. If the farm is cloudy 4 days a week, and clear on three of the remained days, a satellite flying over 7 days a week  will get a good picture at least every week
  • Cloud Removal
    • By combining different images together to create a composite image it is possible to effectively remove clouds from the picture (see example below)
  • Use of Radar
    • Optical satellites cannot see through clouds, but Radar can.
    • Using a radar satellite, such as SENTINEL-1,, which are equired with Radar rathe than optical lenses allow the crops to be analyzed in all weathers, day or night.
    • Synthetic Aperture Radar (SAR) has multiple benefits – one of which is the ability to measure height and detect changes in elevation in millimetre precision.
    • SAR alone would not be sufficient to drive the analysis of crops, but combined optical data it will drive new insights and analysis.

Example:

Cloud Removal Via Composite Images: A satellite passed a location on the three days. On a Monday there is heavy cloud, Tuesday 50% cloud coverage and Wednesday 25% coverage, at best one image has 75% open skies. By collecting a large number of images to make a composite image it will be possible to show 100% of the crop in a single image.

Removing Cloud from Satellite Imagery
Removing Cloud from Satellite Imagery

Scale: Dealing with scale has been a challenge historically – but this is now resolved by the use of satellite which takes wide images ( 90Km+ wide) and yet still maintains a high resolution of 30cm.   Processing this data has also been historically challenging, but this but leveraging modern cloud computing capability means that data can now be collected and processed in a timely fashion

Colour and Farm Size

Trying to identify crops based on the green color, is not effective. I.e. it should not be improved, but not used at all and different methods should be used. Initially, the areas of potential interest, i.e farmland, should be identified and the urban areas discarded and then farmland used for detailed analysis (see below)

Phenology – Determining the Crops

Identification of crops, particularly in a complex environment (small farms, mixed crops, etc) is best achieved via investing in advanced analytics rather than buying more and more data sets of bigger and bigger areas. A method that is known to work is below:

  • Creating a small but accurate labeled data set (human surveyors and drones)
  • Using tried and trusted algorithms to learn what is the required crop.
    • These algorithms can result in 96%+ accuracy for the
  • This accuracy rate can even distinguish between highly similar crops, planted alongside each other such as Sugar and Cassava.

Once the algorithms have been trained to spot the required crop, e.g. Sugar, this can then be automated to calculate the area of crops planted across an entire country, region, or continent.

Area v Yield

The area of planted crops is a key factor and useful to farmers, traders and governments, but it is not as valuable as the yield. Yield, the actual tonnes of grain, sugar or other crop being produced , is correlated to the area of crops and is based on factors such as the health of crops.

Additional data from satellites can provide more details of the crops, including chlorophyll levels, the quality of the crop, risk of disease and even moisture in the earth. This data, together with the projected area and historical information can be used to predict the yield.

Summary:

Predicting the areas of crops being grown is a challenge, but with the correct use of satellites and analytics highly accurate predictions can be obtained.