Assessment of accuracy
The CEH Land Cover® plus (LC+) Crops map was validated using land parcel data collated by Rural Payments Wales and Scottish Government Rural Payments and Services. These datasets comprise a record of all land registered for agricultural use and are submitted annually by farmers. Crops are assigned to unique land parcels held in the using the Land Parcel Identification System (LPIS). We intersected the crop map with LPIS parcels, excluding those where a one-to-one match between the LC+ Crops map polygon and the LPIS parcels was not clear. We also excluded LPIS parcels where multiple crops had been registered in the same year, as the locations of each different crop within the parcel are not recorded and so cannot be compared to corresponding LC+ Crops polygons with multiple crops.
After these exclusions, the final validation dataset consisted of crop data on around 230,000 fields covering 1.1 million hectares of agricultural land in Scotland and Wales. Comparing these with the corresponding LC+ polygons showed a total accuracy across classes of 97% for Wales and 95% for Scotland.
Accuracy per crop can be measured in two ways: the ability of the LC+ Crops map to correctly detect known crops (% detected) and the proportion of crop assignments in LC+ Crops which are correct (% correct). In many cases these two measures of accuracy are similar and accuracy in terms of both detection and correct identification was uniformly high for the UK’s dominant crops: oilseed rape (> 95%), grass (> 96%) and winter cereals (> 84%). Validation with LPIS parcels suggested the combination of winter wheat with winter oats, as this combined class gave greater overall and per class accuracies. The results of the validation for Scotland and Wales combined is shown below:
Full validation tables for all crops, in each nation, are available in the metadata. Some crops are expected to be less accurate than others, for example, the LCM+ Crops map ‘Potatoes’ class only includes main crop potatoes. Early crop potatoes are assigned to the ‘Other’ class. The validation data does not differentiate between these two, so some apparent misclassification is to be expected.
Areas for improvement in future releases
Reduced accuracy in the 2015 release of the LC+ Crops map may arise from several sources, several of which are targets for improvement in the 2016 release:
- The 2015 LC+ Crops map is based on classifying land which falls within the arable and improved grassland classes of the LCM2007. Where these were inaccurate in the original classification, the LC+ Crops map will be missing data or attempting to assign a crop to parcels which are not agricultural. The latter is being resolved by including additional parcels which were classified as rough grassland in LCM2007 and using ordnance survey data to filter out non-agricultural land parcels.
- The combined classes for winter cereals may be split into individual crop classes in future releases, once classification for these crops has been improved.
- We are addressing the remaining confusion of grass with spring cereals this by the inclusion of Sentinel 2 optical data to refine our classification of these crops.