In this post, a validation of the land-cover map of France produced by CESBIO for the 2016 period was presented. This validation used independent data (that is data collected by different teams and using different procedures than the data used for the classifier training), but the validation procedure consisted in applying classical machine learning metrics which, as described in this other post, have some limitations.A fully independent validation following a sound protocol is costly and needs skills and expertise that are very specific. SIRS is a company which is specialised in the production of geographic data from satellite or aerial images. Among other things, they are the producers of Corine Land Cover for France and they are also responsible for quality control and validation of other Copernicus Land products.SIRS has recently performed a validation of the 2016 France land-cover map. The executive summary of the report reads as follows:
This report provides the evaluation results of the CESBIO OSO 2016 10m layer and the CESBIO OSO 2016 20m layer.The thematic accuracy assessment was conducted in a two-stage process:
- An initial blind interpretation in which the validation team did not have knowledge of the product’s thematic classes.
- A plausibility analysis was performed on all sample units in disagreement with the production data to consider the following cases:
- Uncertain code, both producer and operator codes are plausible. Final validation code used is producer code.
- Error from first validation interpretation. Final validation used is producer code
- Error from producer. Final validation code used is from first validation interpretation
- Producer and operator are both wrong. Final Validation code used is a new code from this second interpretation.
Resulting to this two-stage approach, it should be noticed that the plausibility analysis exhibit better results than the blind analysis.The thematic accuracy assessment was carried out over 1,428 sample units covering France and Corsica.The final results show that the CESBIO OSO product meet the usually accepted thematic validation requirement, i.e. 85 % in both blind interpretation and plausibility analysis. Indeed, the overall accuracies obtained are 81.4 +/- 3.68% for the blind analysis and 91.7 +/- 1.25% for the plausibility analysis on the CESBIO OSO 10m layer. The analysis on the 20m layer shows us that the overall accuracy for the blind approach is 81.1 +/-3.65% and 88.2 +/-3.15% for the plausibility approach.Quality checks of the validation points have been made by French experts. It should be noticed that for the blind analysis, the methodology of control was based mostly on Google Earth imagery, no additional thematic source of information that could provide further context was used such as forest stand maps, peatland maps, etc.
These results are very good news for us and for our users. The report also contains interesting recommendations that will help us to improve our algorithms. The full report is available for download.