Comparing the accuracies of remote sensing global burned area products using stratified random sampling and estimation

TLDR: In this article, the accuracies of six global burned area (BA) products for year 2008 were compared using the same validation methods and reference data to quantify accuracy of each product.

About: This article is published in Remote Sensing of Environment.The article was published on 2015-04-01 and is currently open access. It has received 159 citations till now.

Figures

Table 3: Estimated error matrices and reference burned area percent (BA%) for each product at the 250

Table 1: List of products included in the analysis 111

Table 2: Population error matrix where Pij is expressed as proportion of area of agreement (diagonal 156 cells) or disagreement (off diagonal cells) between the BA product (map) class and the reference class 157 (∑Pij=1, and the row and column margins are the sum of the cell entries in that row or column). 158

Figure 2: Accuracy measures of DC on TSAs for MCD64 and MERIS_cci (TSA spatial extents are 302 exaggerated to enhance visualization). 303 304

Figure 3: Accuracy measures of B on TSAs for MCD64 and MERIS_cci (TSA spatial extents are 306 exaggerated to enhance visualization). 307 308

Figure 1: Spatial distribution of the 103 sample TSAs (black polygons) selected by stratified random 130 sampling (for context, % burned area in 2008 at 0.5 o

Frequently asked questions

Q1. What are the contributions in this paper?

In this paper, the accuracy of six global burned area ( BA ) products for year 2008 were compared using the 22 same validation methods and reference data to quantify accuracy of each product. 

Q2. What is the main reason for the comparison of BA products?

348An important feature of the validation and comparison of BA products based on cross-349 tabulation analysis is that the comparison of BA products can be based on accuracy metrics 350 selected to address specific end-user requirements. 

Q3. How many regions were allocated to the 115 probability sampling design?

The 115 probability sampling design employed a spatial stratification to distribute the sample among 116 the major Olson biomes (Olson et al. 2001), with proportionally larger sample sizes allocated to 117 regions with high BA. 

Q4. What was the purpose of the ESA CCI Fire Disturbance 48 project?

The ESA CCI Fire Disturbance 48 project (fire_cci) aimed to develop global burned area products from European sensors for the 493climate modeling community, with proper validation and uncertainty characterization 50 (http://www.esa-fire-cci.org/, last accessed December, 7th 2014). 

Q5. What was the algorithm used to extract the fire perimeters?

For each TSA sampled, fire perimeters were extracted from a pair of 137 Landsat TM/ETM+ image acquisitions at the same location (acquired in two different revisit 138 times at the same path and row), using a semi-automatic algorithm developed by Bastarrika et 139 al. (2011). 

Q6. What was the coded error matrix for the product pixels?

To build the error matrices, product pixels were coded as “burned” if fire 146 was detected between the reference image acquisition dates. 

Q7. What are the metrics used to evaluate the accuracy of the fire_cci products?

users expressed interest in metrics providing 83 estimates of accuracy, commission and omission errors, error bias (whether the product under 84 or overestimates true BA) and temporal stability (covered in Padilla et al. 2014b). 

Q8. What is the significance of differences in accuracy measures among BA products?

Statistical significance of differences in accuracy measures among BA 312 products was evaluated taking into account the probability sampling protocol. 

Q9. What is the reason for the large underestimation of BA extent?

The common large underestimation of BA extent may be caused by the coarse spatial 330 resolution of global products and the large amount of area burned in small fire patch sizes. 

Q10. How many previous studies have been documented?

The sampling design, reference data generation and methodology for estimating accuracy had 114 previously been documented in Padilla et al. (2014a), where further details are included. 

Q11. How many pixels were available for the 105 TSAs?

The proportion of available pixels 27115in the 105 TSAs was near 82% for VGT_cci and MCD64, 97% for MCD45, and 100% for the 272 other products.