ACCURACY ASSESSMENT ANALYSIS OF MAPBIOMAS’ LAND COVER (AND LAND USE) MAPPING

The accuracy estimative for Collection 3.0 is still in preparation

Accuracy analysis is the main way of assessing the quality of mapping performed by Mapbiomas. In addition to telling the overall classification accuracy, the analysis also reveals the accuracy and error rate for each classified class. MapBiomas evaluated global and per-class classification accuracy for each year between 2000 and 2016.

Accuracy estimates were based on the evaluation of a sample of pixels, which we call the reference database. The number of pixels in the reference database was pre-determined by statistical sampling techniques. Each year, each pixel in the reference database was carefully evaluated by three technicians trained in visual interpretation of Lansat images. The evaluation of a pixel in a given year was considered valid only when two or three technicians agreed on the class observed in the pixel. Accuracy assessment is then done using metrics that compare the class mapped with the class evaluated by the technicians in the reference database. We report yearly accuracy estimates with c.a. of 5% error for each class in the mapping.

In each year, the accuracy analysis is done by cross-tabulating the sample frequencies of the mapped and real classes, in the format of Table 1. The frequencies ni,j represent the number of pixels in the sample classified as class i, and evaluated as class j. The line marginal totals, Formula1, represent the number of samples mapped as class i, while the column marginal totals,Formula2 , represent the number of samples that were evaluated by the technicians as class j. Table 1 is commonly called the error matrix or confusion matrix.

Table 1: Generic sample error matrix

Tabela1 mapbiomas en

From the results in Table 1, the sample proportions in each cell of the table are estimated by Formula3 Formula4. The matrix of values Formula5 is used to estimate:

  1. User’s Accuracies: These are the estimates of the fractions of pixels, for each classified class, correctly classified. The user’s accuracy is associated with the error of commission, which is the error of assigning a pixel to class i, when it belongs to some other class. The user’s accuracy for class i is estimated by Formula6 and the commission error by Formula7. These metrics are associated with the reliability of each classified class.
  2. Producer’s Accuracies: Are the sample fraction of pixels of each land cover/use class correctly assigned to their classes by the classifiers. The producer's accuracy is associated with the omission error, which occurs when we fail to map a class j pixel correctly. The producer 's accuracy for class j is estimated by Formula8 and the omission error by Formula9. These metrics are associated with the sensitivity of the classifier, that is, the ability to correctly distinguish one class from another.
  3. Global Accuracy: It is the estimate of the overall hit rate. The estimate is given by Formula10 n, the sum of the main diagonal of the proportions matrix. The complement of the total accuracy, or the total error Formula11 is still decomposed into area (or quantity) disagreement and allocation disagreement1. Area disagreement measures the fraction of the error attributed to the amount of area allocated incorrectly to the classes by the mapping, while the mismatch allocation to the ratio of class-displacement errors.

The matrix also provides estimates of the different types of errors. For example, we show estimates of true class area composition in each mapping class. Thus, in addition to the hit rate of the class mapped as forest, for example, we also estimate the fraction of these areas that could be pasture or other classes of cover and land use, for each year. We understand that this level of transparency informs users and maximizes the potential of use across multiple types of users. For this, we build an application to facilitate the visualization of the accuracy and the errors of the mapping

1- Pontius Jr, R. G., & Millones, M. (2011). Death to Kappa: birth of quantity disagreement and allocation disagreement for accuracy assessment. International Journal of Remote Sensing, 32(15), 4407-4429.


About the graphics


Overall Statistics

They show the mean annual total accuracy and the error, decomposed in area and allocation disagreements.

Acuracia img1 en

Graph 1. Annual Total Accuracy Chart:

This graph shows the total accuracy and the total error per year. The total error is decomposed into area and allocation disagreements. Accuracy is plotted at the top and errors at the bottom of the chart.

Acuracia img2 en

Graph 2: Matrix of errors

This graph shows the user’s and producer’s accuracy, and the confusions between classes for each year. The first shows the confusions for each mapped class. The second shows the confusions for each real class.

Acuracia img3 en
Acuracia img4 en

Graph 3: Class History

This graph allows you to inspect the confusions of a particular class over time. The user’s r and producer’s accuracies for each class is displayed along with the confusions in each year.

Acuracia img5 en
Acuracia img6 en