REDDAFDeforestation Mapping

EO-Based Forest Cover Change Mapping Services have been Improved

The Cameroon REDD Pilot Project demonstrated improvement needs related to Forest Cover Change mapping in the Congo Basin Region. Thus, the following research activities have been realized within REDDAF:

  • Methodologies were developed to combine various multi-temporal and multi-scale optical sensor data with radar sensor data (in order to fill the gaps induced by frequent cloud cover in tropical areas).
  • A robust radiometric calibration technique was developed in order to facilitate classification and to improve the usability of multi-sensor data.
  • Additionally, an approach to radiometrically correct areas affected by cloud shadows was developed to enhance the use of historic image data.

Test Areas in Cameroon & CAR

The methods were developed in an area in the Center Province of Cameroon and subsequently were tested in an area within the Central African Republic. The methods developed have been integrated into optimized and cost efficient operational processing chains.

Integrating Optical and SAR Data

As optical data often is affected by clouds and cloud shadows, alternative approaches are investigated to fill gaps in the optical data coverage. We have investigated the potential for using SAR data for deforestation mapping and a processing chain was implemented for both pre-processing and classification of PALSAR data. For a proper integration of optical and SAR data, geometric congruence of the data sets is a prerequisite. Typically, this can be achieved by time-consuming manual tie point selection, which is difficult in homogeneous forested areas. In order to reduce these manual efforts, a fully automatic image matching procedure using the mutual information method was developed. The results matching PALSAR and Landsat images show that the RMSE could be reduced from over 80 m to less than 10 m in both X and Y direction. Tests have been done with a wide variety of different data sets with similar results. This pre-processing step is important for both deforestation and also for degradation mapping.

REDDAFLeft image: before matching; right image: after matching

Following the successful pre-processing, a method has been developed for the cloud gap filling. The classification-based trainer was developed, which allows filling classification gaps caused by clouds or sensor failures in optical data by using SAR data without much manual effort.


Left image: Forest (purple), nonforest (blue) classification result from optical data with gaps
Right image: Information gaps filled by automated SAR classification using the classification-based trainer

Adjusting Areas Affected by Cloud Shadows

For maximizing the use of optical satellite data, a method was developed to adjust areas affected by cloud shadows. The method is based on morphological characteristics and radiometrically enhances areas covered by cloud shadows (see Figure below). Correction is performed by histogram matching to the surrounding area. The corrected areas may be integrated into classification procedures and thus avoid separate classification or even visual interpretation of these areas, resulting in a time- and cost-saving processing line.

REDDAF - Reducing Emissions from Deforestation and Degradation in Africa














Cloud shadow compensation on Landsat image subset (clouds masked out in black). Left image: Landsat input image (with cloud shadow examples). Right image: Cloud shadow compensation.












Degradation Mapping

Development and Testing of EO Methods

A main objective in the R&D of REDDAF was to develop viable Earth Observation (EO) methods for mapping degradation in the Congo Basin by overcoming the technical restrictions related to:

  • frequent cloud cover in tropical forest areas and
  • complex degradation patterns in tropical forest areas.

The main task was to develop tools and processing chains which allow mapping:

  • the extent of degraded areas and
  • the degree of degradation.

Methods were developed to make use of various sensor data for degradation mapping. An approach based on spectral mixture analysis was enhanced by introducing multi-temporal aspects based on optical data (time series analysis). Furthermore, options for 3D mapping were investigated in order to detect the gaps in the forest canopy induced by logging activities. This 3D approach is based on SAR data.

Test Areas in Cameroon & CAR

The methods for degradation mapping were developed in the Pallisco concession site in South-East Cameroon. In the second step, the methods were transferred to map a prototype area in CAR.

Mapping Degradation Areas by Means of Time Series Analysis

A multi-temporal method for forest degradation mapping caused by selective logging was developed.  Based on visual interpretation of very high resolution optical data as reference, a classification was performed. The time series analysis builds on a multi-temporal stack of spectral mixture analysis (SMA) fraction images. The classifier used is a minimum distance classifier. For the training, the temporal curves from the reference data were used. The pixel-based result of this classification (gaps, roads, skid trails, etc) was then refined by context analysis (distance to logging roads / settlements). Finally, the classified pixels were aggregated to larger areas of potential forest degradation with a minimum mapping unit of 5 ha.  An example mapping result for a small subset of the prototype area in CAR is shown in the Figure below.

Left image: Landsat image 2001 in CAR.
Right image: Mapping result (non-forest areas in white and degradation areas in red) superimposed on Landsat image 2001.

3D Mapping of Forest Canopy to Detect Degradation

Furthermore options for 3D mapping of forest canopy to detect degradation have been investigated. The work comprised InSAR processing, radargrammetric processing, and the extraction of degradation areas from digital surface models. Results show that from CSK data a 3D model could be derived, which in combination with the existing SRTM model clearly shows gaps and roads as features of degradation (dark patches in the right Figure below).These features were then aggregated to areas of degradation. The accuracy assessment of this aggregated result compared to visual interpretation showed an overall accuracy of 82%.

Left image: Test area for 3D degradation mapping, Rapideye data from December 2011. Red: areas affected by degradation; blue: extent of 3D model from CSK data.
Right image: Difference model from CSK 3D generated with the SRTM difference approach.

Direct Biomass Assessment

REDDAFDirect Biomass Assessment

EO-Based Methodologies to Directly Assess Above-Ground-Biomass

The objective of the biomass study was to develop, test and provide improved methodologies to directly assess above-ground-biomass in the Congo Basin Region using radar EO data. The activities aimed to develop transferrable methodology to provide mapping products that contain gridded and georeferenced values of Above Ground Biomass (AGB) and related uncertainties in AGB. The approach focused on improving existing image processing and biomass inversion methods to fully exploit the potential of the currently available data (ALOS PALSAR-L-band). The work that was based on existing efforts conducted by CESBIO, has been tested and refined using in situ data collected in Cameroon.

Improving the State-of-the-Art Methods

Methods beyond the state of the art in biomass mapping in Cameroon were assessed, for improvement in several aspects:

  • Development of methods which can be generalized on other sites, and relying on a limited number of in-situ data.
  • Improvement of the spatial resolution of the biomass map by using multi-temporal and dual polarized ALOS PALSAR data for speckle filtering without degrading the spatial resolution.
  • Improvement of the inversion results using Bayesian inversion and/or the Support Vector Regression (SVR) method. 
  • Development of a new method for uncertainties assessment.

The processing chain was achieved mostly using free software for practical implementation. The methods were applied to ALOS-PALSAR data of the prototype region (about 112 x 112 km) to provide biomass maps for years 2007-2008-2009-2010.  The methods were further validated in another region in Cameroon.

Research Areas in Cameroon

The prototype region (Figure 1, no.1) is located in Adamawa region, central Cameroon, centered around Mbakaou lake near the departmental capital Tibati, encompassing the Mbam Djerem National Park. This region was chosen as it extends across a range of tropical vegetation types, from humid forests contiguous with the Congo Basin tropical forest belt in the south to savanna with narrow gallery forests in the north.

The validation site was planned to be conducted in CAR. Because of the security conditions in CAR, it was however decided to conduct field campaign in another region of Cameroon, near Bafia (Figure 1, no.2). This region holds large and homogeneous forest areas and a large range of biomass.

REDDAFFigure 1: GlobCover 2009 classification map over Cameroon. The first test site (Adamawa Province) is indicated by the black rectangle 1 and the second test site (Central Province) is indicated by the black rectangle 2.

Biomass Map for Prototype Region Completed

The inverse model relating radar backscatter to AGB was applied to the whole prototype region. Figure 2 shows the resulting biomass map. The map is composed of a class of water and digital values of biomass (displayed here in six classes). This biomass map highlights the dense humid forests (Mbam and Djerem National Park) with biomass > 150 ton/ha, and the gallery forests in the savanna, with biomass lower than 100 ton/ha. It should be noted that large areas of gallery forests and transitional forests (with biomass reaching 60-80 ton/ha) contain a large amount of carbon stocks, which is often neglected in carbon estimates.

REDDAF - Reducing Emissions from Deforestation and Degradation in Africa

Figure 2: Biomass map of the region of Adamawa (about 112 km x 112 km), central Cameroon. The biomass value for each 25 m pixel is in ton/ha and was derived from 4 FBD ALOS-PALSAR data, acquired on June 18, 2010 (East track) and July 5, 2010 (West track). ALOS PALSAR data acquired in 2007, 2008, and 2009 have been used in addition for speckle filtering, to enhance the data radiometric quality and the product spatial resolution.

Figure 3 - extract of Figure 2: Biomass map of the 20 km x 20 km region North of Mbam and Djerem National Park, centered at 5° 57’17’’ N and 12° 46’ 08’’ E

Figure 4 (below. left): Validation of inversion results using the inverse model based on in situ plots measured in the prototype  site (Adamawa), calibrated using three in situ plots measured at the validation site (Bafia). The RMSE is 11.5 t.ha-1. (below, right): The most difficult tree to measure.

REDDAFFor method validation, the model relating radar backscatter to in-situ AGB developed at the Adamawa site was applied to the Bafia site with very good agreement. For AGB lower than 150 t/ha, a significant correlation is obtained when relating radar backscatter and in situ AGB with an rp of 0.86 and a p value of 0.8x10-4. The inverse model has been applied to the 2010 PALSAR mosaic data over the whole Bafia site, and the in situ data measured in this site were used to assess the accuracy of the AGB map. A calibration of the model was required with a small number of in situ plots (3 to 5) to account for REDDAFthe local conditions (forest types, environmental conditions). The result shows that using three in situ plots (e.g. AGB of 22.1, 47.2 and 112.3 t.ha-1), the agreement is excellent with a RMSE of 11.5 t.ha-1. Figure 4 shows the validation of inversion results restricted to retrieved AGB below 150 t/ha. The method can be therefore considered fully validated. The proposed processing chain is viable and can be implemented for the whole of Cameroon.

Figure 5a shows the resulting AGB map of Bafia (112km×112km, or 1 degree square from 4°N to 5°N and from 11°E to 12°E). An uncertainty map is associated with the AGB map (Figure 5b). Figure 6 shows a 522 sqkm window from the map in Figure 5. The biomass map is composed of digital values of biomass (displayed here in sixteen classes, from 0 to 150 t.ha-1 and more). The results have been used to compute the biomass over the whole Bafia region. There are 82.98 Mt of biomass or 41.49 Mt of carbon in approximately 12 550 sqkm, representing  a mean AGB of 66.12 t.ha-1 The map of uncertainties is composed of digital values of biomass displayed in fifteen classes. The mean value of uncertainties is about 11 tha-1. 

REDDAFFigure 5: a) Biomass map (AGB) of the region of Bafia, Central Cameroon (about 112km×112km, or 1 degree square from 4°N to 5°N and from 11°E to 12°E), and b) AGB uncertainties map of Bafia. The values for each 25 m pixel are in ton/haand were derived from  ALOS-PALSAR mosaic data, acquired on 2010. ALOS PALSAR data acquired in 2007, 2008, and 2009 have been used in addition for speckle filtering, to enhance the data radiometric quality and the product spatial resolution

REDDAFFigure 6 - Extract of Figure 5: a) Biomass estimates (AGB) int.ha-1, b) AGB uncertainty estimates (SD(AGB)) in t.ha-1 (same colour codes as Figure 5) and c) optical image (from Google Earth) over a 522 km2 area centred on 4.59°N and 11.77°E.

Further Research Needed at Different Levels

  • The refinement of the inversion algorithm aiming at estimating biomass at a refined spatial resolution  (e.g. 15 m) can be continued in further research activities.
  • The stocks of carbon in the savanna-forest ecotone have to be quantified and their changes between 2007-2010 have to be also assessed. 
  • The method should be applied and assessed in Central African Republic.