Improving the model
The published results from the Scottish Crop Map 2019 come from the first iteration of the machine learning model. Work will continue to improve the model in light of the limitations discussed in this paper.
The next step is to start including other data sources into the model to provide different types of signal for each crop. This should help to improve the predictions for some of the crops we have already included in the model but allow us to include smaller crop types which couldn't be included in this first iteration.
There are a number of data sources we will start to incorporate soon into the model. This includes a day of cloud-free Sentinel-2 satellite images for Scotland and weather variables (including hours of sunlight, average temperature and rainfall).
More timely crop predictions
Looking forward we aim to produce predictions in a more timely manner. To support us in this we have produced a guidance document and worked to streamline our methods. We have also carried out preliminary testing of the current model to identify whether all months of radar images are necessary. Early findings suggest the Autumn months are not required for our current predictions.
Estimating production values and yields
Estimating production values and yields is an important step in the development of this project, as these could become more timely and granular than what can be supplied through the current method. We are looking to collaborate with other organisations on this work and include data such as soil and weather variables.
Communicating with the team
Please contact us to query any of our work or to discuss collaboration projects. You can contact us at firstname.lastname@example.org.
|CEDA||(The Natural Environment Research Council's Data Repository for Atmospheric Science and Earth Observation) Centre for Environmental Data Analysis|
|Grassland||Only permanent grassland is included in our model|
|Ground truth||Information (data) collected from the location and therefore are known.|
|JASMINE||JNCC's data analysis virtual machine available through the Simple ARD service.|
|JNCC||Joint Nature Conservation Committee|
|LPIS||Land Parcel Information Service, part of Scotland's rural payments and pervices|
|Mask||A dataset that defines which locations/ objects will be excluded from a geospatial file|
|Neural network||A machine learning algorithm that attempts to learn from data in a way that mimics the human brain|
|Polarisation||The orientation of the radar signal. Radars can transmit and receive horizontally polarized (H) or vertically polarized (V) signals|
|Polygon vector||A coordinate-based geographic data model that uses points, lines and shapes to represent features|
|Producer accuracy||The number of fields classified as a particular crop type by the model, as a proportion of the number of fields known to be that crop|
|Random forest||A machine learning algorithm based on using data to make repeated decisions that lead to an optimal classification|
|RFI||Radio frequency interference|
|SAF||Single Application Form, the form used to claim payments under a number of agricultural support schemes|
|Supervised learning||Any of a class of machine learning algorithms that use existing knowledge of categories or values to train a model to make new decisions using the same categories or values|
|Zonal statistics||Summary statistics about individual spatial areas|