Landsat-based reference data

We used Landsat-based reference data from five NELDA test sites: Amur (AMUR), Chita (Chita), Sikhote-Alin (SIKH), St. Petersburg (STPB), and Vasyugan (VASY). We aggregated several classes from the original Landsat-based change maps to represent forest disturbances, forest regrowth, undisturbed forest, and non-forest land cover for the period between 2000 and 2006 (Table 1). To determine the initial forest area, we used Landsat-based land cover maps from ca. 2000. All pixels outside this initial forest mask were reclassified to non-forest land cover. The resulting forest change (2000-2006) maps were then used as reference data in the subsequent analyses.

AMUR CHITA

SIKH STPB

VASY

VASY

We calibrated and validated the MODTrendr map at the MODIS pixel level, i.e. the MODIS pixels served as sampling units and the Landsat-based disturbance maps were aggregated to the MODIS pixel size 463.3 m x 463.3 m). We only included pure forest samples with an initial forest area proportion of 0.95 and greater based on the proportion of forest pixels derived from the ~2000 land cover maps. Only samples with 100% cloud-free Landsat change observations were included.

For each sample, we estimated the reference proportion of forest disturbance by calculating the proportion of Landsat-based forest disturbance pixels within each MODIS pixel. The proportion of total Landsat-based forest area included in sample pixels ranged from 69% in SIKH and 53% in CHIT to 38% in AMUR and 23% in VASY, where forest cover was very fragmented. Finally, the pure forest sample was randomly split (stratified by disturbed and undisturbed forest) into a calibration set (40%) and a test set (60%).

Disturbance Proportion

Distribution of disturbed area proportion from Landsat for pure forested MODIS pixels at NELDA test sites

Credit:

  • D. Pflugmacher, D. Sulla-Menashe, O. Krankina. “ASSESSMENT OF THE MODTRENDR ALGORITHM FOR MAPPING FOREST DISTURBANCES IN NORTHERN EURASIA” NELDA-II REPORT (UNPUBLISHED)