Generation and comprehensive validation of 30 m conterminous United States Landsat percent tree cover and forest cover loss annual products

Abstract

This study describes the generation and comprehensive validation of 30 m Landsat-based annual percent tree cover and forest cover loss products for the conterminous United States (CONUS). The products define (i) forest status with respect to three thematic classes: stable forest, stable non-forest, forest cover loss, (ii) percent tree cover (PTC, 0–100%), (iii) percent tree cover decrease (ΔPTC), and (iv) the Landsat acquisition dates bounding mapped forest cover loss occurrence. Forest was defined, based on the U.S. federal government forested land definition, as 30 m pixels with mapped PTC >10%. Annual products were derived using temporally overlapping 9-year periods (mapping within each central 5-year period) of USGS Landsat Analysis Ready Data (ARD) with reconciliation of the results between periods. The products for 2013 are presented and were validated rigorously by comparison with 1910 30 m independent reference data interpreted from bi-temporal <1 m resolution aerial imagery selected using a Stage 3 CONUS stratified random sampling design. The stable forest, stable non-forest, and forest cover loss results were validated using standard accuracy metrics derived from the confusion matrix. The overall accuracy was high (0.92), and class-specific user's accuracy (UA) and producer's accuracy (PA) metrics were also high for the stable forest (UA = 0.94, PA = 0.84) and stable non-forest (UA = 0.90, PA = 0.97) classes. The forest cover loss class had similarly high UA (0.89) but significantly lower PA (0.61) indicating non-negligible omission errors. All standard errors were <5%. The total area of stable forest over CONUS for year 2013 was estimated as 3,049,380 ± 114,392 km2 and the total area of forest cover loss was estimated as 31,382 ± 4751 km2, with 95% confidence interval. The PTC and ΔPTC products were validated by linear regression with the reference data, indicating good PTC precision reflected by a high coefficient of determination (R2 = 0.79), and accuracy with a regression slope close to unity (0.86) and small intercept (3.48). The regression between mapped ΔPTC and the reference data had a high coefficient of determination (R2 = 0.74) but a regression slope further away from unity (0.78) and small intercept (1.68) consistent with the forest cover loss omission errors revealed by the confusion matrix. State-level comparison of the stable forest mapped area with forest land area statistics published by the U.S. federal government for the 48 CONUS states indicated reasonable correspondence (R2 = 0.97) but with a 1.15 regression line slope indicating relative over estimation of the mapped stable forest area, likely related to forest land reporting differences.