A goal of large-area land cover mapping is to produce globally consistent characterizations that have local relevance and utility; that is, reliable information across scales. (Hansen et al 2013)
This package allows to obtain yearly forest cover maps from the Global Forest Cover dataset by Hansen et al (2013) and run a pixelwise comparison with Colombian Forest Cover Datasets produced by the Forests and Carbon Monitoring System - SMBYC at the National Institute for Meteorology and Environmental Studies (IDEAM).
You can install the development version of LCoverFlow from GitHub with:
# install.packages("devtools")
devtools::install_github("RSENSUS/harmonizacion_hansenIdeam")- Assess the level of agreement between forest cover products by iteratively comparing multiple canopy cover for GLAD forest products with the national forest cover assessment
- Identify the threshold that produces the highest level of overall agreement between both datasets
- Perform the process for ecologically homogeneous spatial units.
-
Fill spatial and temporal gaps in national forest coverage reporting, but homogenized to the national standards. Reduce costs and improve update frequency .
-
Use local experience and locally produced datasets as validation data.
-
Identify assess over/sub register of forest coverage/forest loss in the national and global datasets.
Continental extent of Colombia, divided in 379 spatial units (biomes). (Londoño et al, 2017)
National data on forest cover and change for Colombia (IDEAM) The Forest and Carbon Monitoring System (Sistema de Monitoreo de Bosques y Carbono) ascribed to IDEAM in Colombia (Table 1) produces the official datasets on forest cover and change for the country and is the reference for national and regional policy (DNP, 2020) and to fulfill Colombia’s international reporting obligations (IDEAM, 2019). The forest cover product represents forest presence in areas where vegetation meets the national definition of forests (Table 2). This dataset explicitly excludes commercial forestry, oil palm, and trees intended for agricultural use (Galindo et al., 2014a,b).
Forest change is detected through a Principal Component Analysis of merged data of end and start years. Forest change maps are produced first by overlaying the maps representing forest cover as described above. These maps then are subjected to a quality control process that consists of manual reclassification of mistaken pixels based on visual interpretation of the last available images for the year (IDEAM, 2019). Any pixels with low quality observations due mostly to the presence of clouds or cloud shadows in at least one of the mapped years are removed from the analysis of forest loss in the entire time series. Maps were produced at irregular intervals for the years 1990, 2000, 2005 and 2010 and annually from 2012 onwards. The IDEAM national forest change data is the reference source to produce the official national report of annual rates of forest loss and therefore it was used for comparing estimates of forest cover loss with the GFC product. For change between 2012-2013 we used V5, for 2013-2014 we used V6, for 2014-2015, 2015-2016 , and 2016-2017 we used V7, and for 2017-2018, and 2018-19 we used V8. We used Magna-Sirgas as the coordinate system for map comparison, given that this is the official reference coordinate system for Colombia.
The High-Resolution Global Maps of 21st-Century Forest Cover Change (GFC- Table 1) is produced by the Group on Land Analysis and Discovery (GLAD) at the University of Maryland (Hansen et al., 2013). GFC harmonizes all available spectral data from the Landsat mission to derive descriptive annual and semi-annual statistics (median, maximum, minimum, and quantiles) for each spectral band that constitutes data inputs to produce annual forest loss estimates through the application of a supervised decision tree classifier. The final GFC product is constituted by two main data sets. Pixel values for the first one represent the percentage canopy cover in the first year of analysis (2000). The second layer, assigns a value representing the year of forest cover loss to all pixels that were mapped with a tree cover higher than 30% in the reference year (2000). This layer enables users to define forest extent in terms of a minimum percent tree cover threshold. This threshold can be used to produce a mask of forest cover for the initial year with a tree cover above the user-specified threshold. This mask can then be applied to the layer on forest loss year to remove pixels labeled as loss that are under the defined tree cover threshold (Hansen et al 2013).
This table provides a comparison between the GLAD and SMBYC datasets used for land cover analysis.
| Criteria | IDEAM | GLAD |
|---|---|---|
| Spatial Resolution | 30m | 30m |
| Base Information | Landsat | Landsat |
| Coverage | National | Global |
| Frequency | Irregular | Annual |
| Years | 1990, 2000, 2005, 2010, 2012, 2013, 2014, 2015, 2016, 2017, 2018, 2019 | 2000-2020 |
| Strengths | - Distinguishes between planted and natural forests - Official information used to measure and report forest cover in Colombia |
- Higher temporal resolution - Less data loss - Annual updates - Multiple canopy cover thresholds available |
| Weaknesses | - Maps produced at irregular intervals - Multiple sources of data loss (SLC-off and cloudiness) |
- Does not distinguish between planted and natural forests - Does not report forest gain |
Harmonization consisted of the production of forest cover maps from GFC that best match the representation of forest cover and change by IDEAM (Fig. 1) and that we refer here as HGFC. The first step was to exclude from the class forest in GFC, pixels representing canopy forming cultivation such as oil palm cultivation and agroforestry systems that are explicitly excluded from the definition of forest used for IDEAM. We eliminated those pixels based on areas representing canopy forming cultivation in the national land cover maps for 2012 and 2017 (IDEAM, 2010 - Table 1), which are the closest dates to the period of analysis.
The second step was to produce a map representing three categories: forest persistence, non forest persistence and forest loss between 2010 and 2017 using the forest non-forest maps produced by IDEAM for those years. We selected 2010 as the first year because it is when the IDEAM maps started to be produced at a higher temporal resolution. We selected 2017 as the last year because that was the latest year available for IDEAM maps at the start of this research. We produced analogous maps from GFC assuming different minimum thresholds in percent tree cover that define forest and non-forest in GFC and used them to identify the threshold that maximized agreement with the map produced from IDEAM. Thresholds ranged between 20 and 100%. Thresholds increased by 10% for values between 20% and 50%, 5% between 50% and 90%, and 1% between 90% and 100%. Finer increments at higher percent tree cover values ensure precision in these ranges, where most of the variation in percent tree cover occurs in the GFC dataset. Forest gain was not considered in the analysis because this data is not available for most years in the GFC product.
To account for possible spatial variations in optimal percent tree cover thresholds within the country, we performed the above described analysis independently for each of the 396 geographic subunits representing areas with homogeneous ecological conditions (Table 1). Data downloading and thresholding was produced using the R package ‘ecochange’ (Lara et. al, 2022). For each subunit, we calculated overall agreement between the change map calculated from IDEAM and each thresholded forest change map calculated from GFC and selected the one that produced the maximum overall agreement for further analysis. We used the optimum percent tree cover thresholds identified for each subunit to derive harmonized maps representing forest cover for the years 2012 to 2021 for the entire country. These maps were used to evaluate class agreement between the maps and to compare estimates of annual forest loss derived from HGFC and IDEAM. Overall class agreement for the entire country was calculated as the area weighted average of the agreements obtained in all subnational units.
## 2. Obtention of yearly harmonized forest
cover maps
The optimized canopy cover threshold for each biome are then used to obtain yearly forest cover maps individually, using the the echanges function again. The procedure has been optimized to run iteratively over the all the individual spatial units. Memory load issues require to divide the download process into smaller subsets and temporarily store the outputs. Next step include reprojecting and aligning the outputs to a common src and assembling the final yearly maps.
This work was supported by The National Aeronautics and Space Administration - NASA, Award 80NSSC18K0339.
Lara, W., Londoño, M.C., Gonzalez, I. and Gutierrez‐Velez, V.H., 2022. ecochange: An R‐package to derive ecosystem change indicators from freely available earth observation products. Methods in Ecology and Evolution, 13(11), pp.2379-2388.
Gutierrez-Velez VH, Rodriguez-Escobar J, Mejía A†, Espejo J, Anaya JA, Blair, ME. Under review. Mapping forest cover and change as continuous variables is essential to advance consistency across forest monitoring products.
Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S. A., Tyukavina, A., Thau, D., Stehman, S. V., Goetz, S. J., Loveland, T. R., Kommareddy, A., Egorov, A., Chini, L., Justice, C. O., & Townshend, J. R. G. (2013). High-Resolution Global Maps of 21st-Century Forest Cover Change. Science, 342(6160), 850–853. https://doi.org/10.1126/science.1244693
Galindo, G., Espejo, O. J., Ramírez, J. P., Forero, C., Valbuena, C. A., Rubiano, J. C., Lozano, R. H., Vargas, K. M., Palacios, A., Palacios, S., Franco, C. A., Granados, E. I., Vergara, L. K., & Cabrera, E. (2014). Memoria Técnica. Cuantificación de la Superficie de Bosque Natural y Deforestación a Nivel Nacional Actualización Periodo 2012 – 2013 [Technical Report]. IDEAM.