GABAM2010 accuracy assessment using stratified random sampling

Abstract

Burned area is one of the key parameters for global and regional carbon cycle and climate change research. Satellite remote sensing technologies provide an effective way for obtaining the large-scale spatial distributions of burned areas. In 2018, the Chinese Academy of Sciences released the first GABAM (Global Annual Burned Area Map) based on Landsat series satellite data. The accuracy assessment of remote sensing data products is of great significance for product users. To date, the accuracy of GABAM products is yet to be independently evaluated. In this study, an accuracy validation study was performed on the 2010 global 30 m spatial resolution burned area product (GABAM2010) for the systematic evaluation of GABAM product accuracy. The GABAM2010 product was validated, and accuracy measures were estimated at the global scale and in several terrestrial biomes, and the technical framework for the accuracy assessment of global remote sensing thematic products was explored. Global accuracy was estimated by stratified random sampling and estimation in a weighted manner on the basis of the error matrix of each sampling unit. This method enables the verification of a large number of products, such as the global burned area product. Stratified random sampling was used in selecting 80 non-overlapping TSA (Thiessen Scene Areas), and reference fire perimeters were determined from multitemporal Landsat TM images for each sampled TSA. Error matrices and six accuracy measures were used in satisfying the criteria specified by the end-users of burned area products. The global validation result showed that the overall accuracy of GABAM2010 was 97.85%, and the commission and omission errors were 24.32% and 31.60%, respectively. The extent of a burned area is often underestimated due to the impact of data quality, such as strips, and clouds. The GABAM2010 products showed high precision in biomes, including tropical and subtropical grasslands, which are highly prone to fire. The high-density burned areas within biomes had higher accuracy than low-density ones. A statistically rigorous accuracy verification method, stratified random sampling was used in verifying the GABAM 2010 product. The method is applicable for the accuracy verification of similar remote sensing products, and the results of the evaluation performed on the GABAM 2010 product illustrated the specific applications of sampling design and accuracy analysis. Compared with simple estimatior, combined ratio estimator with stratified random sampling increased the reliability of the accuracy measures. Simple estimatior is ineffective in resolving some problems arising from the small proportion of an area burned and small number of sampling areas, which result in unreliable accuracy estimates. By contrast, stratification ratio estimation uses global ecological and fire behavior data and can obtain reliable accuracy estimate results.