Confusion Matrix – Values can be entered into the confusion matrix four ways:
- By clicking on one of the "Load Example" buttons at the top of the screen
- By manually typing values into the cells
- Copy / Paste a selection of values (without headers) directly from an Excel document, or
- By importing a matrix as a .csv file using the drag and drop area below the matrix; controls are available to indicate whether the file contains row/header columns and what character is used as a separator if the default settings do not work.
If you wish to delete a class from the matrix, drag the row label of that class to the red box labeled "Drop classes here to delete". Note that this deletion will automatically propagate through the merged matrix and all computed metrics.
Merged Matrix – Categories in the original matrix can be merged by clicking on the "Merge classes" button. This will bring up a dialog box with blue boxes containing the original categories on the left, and a column of empty grey boxes to the right. The blue boxes can be dragged into the grey boxes; all categories that you wish to combine should be dragged into a single grey box. When you are finished, click on the "Close" button, and two new features will appear on the screen: a "Merged Matrix" beside the original confusion matrix, and a "Merged Metrics" box beside the overall.
Area-based weighting – Below the confusion matrix is a check box labeled "Calculate weights automatically". It is checked by default. If you un-check it, you can manually edit the values in the "Weights" row of the confusion matrix. These values are the relative abundance of classes on a map and used to correct for stratified sampling designs. The values are internally adjusted so they sum to 1; this means that you can enter the total area of each class on your map, or the relative proportion of each mapped class.
Metrics calculated – References for all of the calculated metrics can be found at the end of the R code found at the link at the top of the page. Confidence intervals (CIs) are ±95% central intervals widthc computed as 1.96*sqrt(variance). As an example, if the overall accuracy of your map is 80%, and the CI is ±6%, then the 95% CI for overall accuracy is between 74% and 86%. If you need variance values, you can either back-calculate them, or use the underlying R code which is recommended to avoid propagating rounding errors.
Downloading results – Results of all the calculations seen on the website can be downloaded by clicking on the "Download Confusion Matrix" button.
Additional features – Some features not available on the website can be accessed in our R code, available at the link near the top of the web page. These include direct calculation of variances for many of the metrics, and functions to compute the statistical significance of differences between confusion matrices for certain metrics.
Citation – If this tool has been useful in preparing a manuscript for publication, please cite our accompanying paper: Salk, C, C Dresel, S Fritz, L See, I McCallum (in review). An exploration of some pitfalls of map accuracy assessment using the new Map Tools resource. Remote Sensing.