*2.2.2.2 Step 8: visual verification of the changes detected*

Once the PCA process has been executed by scene or set of scenes, each interpreter codes each pixel to the corresponding thematic class, obtaining a preliminary forest change map with the following classes or strata:


For this step, each interpreter uses the following information: (i) the last images of each reference year, (ii) the annual composite images for each reference year, and (iii) the median compounds for the last quarter of the reference year. By implementing this step, we seek to evaluate and revise the first version of the map of forest cover change in order to identify inconsistencies with respect to the changes detected in previous years and to generate an adjusted version of the map.

## *2.2.2.3 Step 9: quality control*

The quality control process involves monitoring all the technical steps executed, from the download of the satellite images and the intermediate products to the final results of the Forest Change Map and Forest Cover Map. The SMByC has consolidated a set of tools to guarantee the quality, completeness, and consistency of the data, through a Python script executed in ArcGIS© to produce quality control reports for each scene.

### *2.2.3 Phase 3: thematic accuracy assessment*

Thematic accuracy assessment allows generating reliability metrics of the official forest monitoring data generated to: (1) avoid bias, which means to avoid a systematic underestimate or overestimate of forest cover and forest change, and (2) to reduce uncertainty as much as possible [24], in line with guidelines proposed by the Global Forest Observation Initiative (GFOI). This procedure applies user's precision (commission error) and the producer's precision (omission error) following recommendations developed by [25] over the forest change map. For the execution of thematic accuracy assessment, a team of four experts is formed under the following structure:

• A leader of the evaluation, in charge of coordinating the interpreters, who designs and implements a probabilistic sample. The leader performs the consolidation and verification of the interpretation as well as the accuracy analysis.

**41**

*Colombian Forest Monitoring System: Assessing Deforestation in an Environmental Complex…*

• Three interpreters, with extensive experience in visual and digital interpretation, who are trained to apply the forest definition in medium-resolution optical images.

This analysis consists of the implementation of a stratified random sampling. The size of the sample (*n*) is selected following [26]. The assigned proportion of each class uses a proportional allocation based on the area of each stratum compared to the total area allocated [27]. The proportions of each of the strata are based on the forest change map for the reference year, which are as follows: (1) stable forest, (2) stable nonforest, and (3) deforestation. Additionally, each stratum is subdivided based on a deforestation risk map that identifies two main areas (low

Equation (1) shows the math expression used to calculate the proportion (*Wi*) of mapped area (*Am*,*i*) for each stratum *i*, related to the total area of all classes (*Atot*):

> *Am*,*<sup>i</sup> Atot*

The classes with the largest mapped area are assigned a p-value of 0.9, considering that these classes have a high accuracy, while the change classes, high-risk deforestation areas (DEF-AR) and low-risk deforestation areas (DEF-BR), are assigned a *p*-value of 0.8, considering that for these classes, there is a greater uncertainty and

The standard error (*Si*) associated to each class is calculated as the square root of the variance. The total size of the sample (*n*) is calculated as the sum of the products

(*Si*) of each class, and divided by a general standard error of the classification (*So*),

∑ *i*=1 *n*

For each reference year, a value of 0.005 is assumed for the expected general standard error of classification. The assigned proportion of each stratum is based on a simplified approach to the optimum, based on the proportion of area of each stratum compared to the total area allocated. Thus, the smallest strata are adjusted by minimizing the variance estimator for the accuracy of these user classes, in

The sampling implementation, as well as the sampling point interpretation, is done in Acatama, a self-developed software that is available for QGIS at https://

The interpreter team performs a visual interpretation of each unit of verification (point), applying the forest and deforestation definitions adopted by the SMByC. This procedure is applied using Acatama tool [14] that allows defining a fixed spatial reference scale for interpretation, using all the satellite data available.

(*WiSi*) \_\_\_\_\_\_\_\_ *So* ]

), associated to each class *i*, multiplied by the standard errors

2

(1)

(2)

risk and high risk) based on the historical trends of deforestation.

*Wi* = \_\_\_\_

*n* = [

accordance with the recommendations of [25].

*2.2.3.2 Step 11: sampling implementation*

bitbucket.org/smbyc/qgisplugin-acatama.

*2.2.3.3 Step 12: sampling points interpretation*

*DOI: http://dx.doi.org/10.5772/intechopen.86143*

*2.2.3.1 Step 10: sampling design*

therefore less accuracy.

of the area ratio (*Wi*

squared Eq. (2)

*Colombian Forest Monitoring System: Assessing Deforestation in an Environmental Complex… DOI: http://dx.doi.org/10.5772/intechopen.86143*

• Three interpreters, with extensive experience in visual and digital interpretation, who are trained to apply the forest definition in medium-resolution optical images.

#### *2.2.3.1 Step 10: sampling design*

*Forest Degradation Around the World*

1.Stable forest

2.Deforestation

3.No information

5.Stable Nonforest.

*2.2.2.3 Step 9: quality control*

the following structure:

*2.2.3 Phase 3: thematic accuracy assessment*

4.Regeneration

*2.2.2.2 Step 8: visual verification of the changes detected*

forest change map with the following classes or strata:

reclassification for each class are as follows: (1) stable forest, (2) stable nonforest, (3) deforestation, (4) regeneration, and (5) without information (corresponding to masked data). To adjust the areas without information detected for each reporting period, a time series analysis is applied to verify the temporal consistency. For this process, the information from the most recent reporting period is considered, and the areas "without information" are adjusted compared with the other reporting periods.

Once the PCA process has been executed by scene or set of scenes, each interpreter codes each pixel to the corresponding thematic class, obtaining a preliminary

For this step, each interpreter uses the following information: (i) the last images of each reference year, (ii) the annual composite images for each reference year, and (iii) the median compounds for the last quarter of the reference year. By implementing this step, we seek to evaluate and revise the first version of the map of forest cover change in order to identify inconsistencies with respect to the changes

The quality control process involves monitoring all the technical steps executed, from the download of the satellite images and the intermediate products to the final results of the Forest Change Map and Forest Cover Map. The SMByC has consolidated a set of tools to guarantee the quality, completeness, and consistency of the data, through a Python script executed in ArcGIS© to produce quality control reports for each scene.

Thematic accuracy assessment allows generating reliability metrics of the official forest monitoring data generated to: (1) avoid bias, which means to avoid a systematic underestimate or overestimate of forest cover and forest change, and (2) to reduce uncertainty as much as possible [24], in line with guidelines proposed by the Global Forest Observation Initiative (GFOI). This procedure applies user's precision (commission error) and the producer's precision (omission error) following recommendations developed by [25] over the forest change map. For the execution of thematic accuracy assessment, a team of four experts is formed under

• A leader of the evaluation, in charge of coordinating the interpreters, who

designs and implements a probabilistic sample. The leader performs the consolidation and verification of the interpretation as well as the accuracy analysis.

detected in previous years and to generate an adjusted version of the map.

**40**

This analysis consists of the implementation of a stratified random sampling. The size of the sample (*n*) is selected following [26]. The assigned proportion of each class uses a proportional allocation based on the area of each stratum compared to the total area allocated [27]. The proportions of each of the strata are based on the forest change map for the reference year, which are as follows: (1) stable forest, (2) stable nonforest, and (3) deforestation. Additionally, each stratum is subdivided based on a deforestation risk map that identifies two main areas (low risk and high risk) based on the historical trends of deforestation.

Equation (1) shows the math expression used to calculate the proportion (*Wi*) of mapped area (*Am*,*i*) for each stratum *i*, related to the total area of all classes (*Atot*):

$$\mathbf{W}\_{i} = \frac{A\_{m,i}}{A\_{\text{tot}}} \tag{1}$$

The classes with the largest mapped area are assigned a p-value of 0.9, considering that these classes have a high accuracy, while the change classes, high-risk deforestation areas (DEF-AR) and low-risk deforestation areas (DEF-BR), are assigned a *p*-value of 0.8, considering that for these classes, there is a greater uncertainty and therefore less accuracy.

The standard error (*Si*) associated to each class is calculated as the square root of the variance. The total size of the sample (*n*) is calculated as the sum of the products of the area ratio (*Wi* ), associated to each class *i*, multiplied by the standard errors (*Si*) of each class, and divided by a general standard error of the classification (*So*), squared Eq. (2)

$$m = \left[\frac{\sum\_{l=1}^{n} (W\_l S\_l)}{S\_o}\right]^2\tag{2}$$

For each reference year, a value of 0.005 is assumed for the expected general standard error of classification. The assigned proportion of each stratum is based on a simplified approach to the optimum, based on the proportion of area of each stratum compared to the total area allocated. Thus, the smallest strata are adjusted by minimizing the variance estimator for the accuracy of these user classes, in accordance with the recommendations of [25].

#### *2.2.3.2 Step 11: sampling implementation*

The sampling implementation, as well as the sampling point interpretation, is done in Acatama, a self-developed software that is available for QGIS at https:// bitbucket.org/smbyc/qgisplugin-acatama.

#### *2.2.3.3 Step 12: sampling points interpretation*

The interpreter team performs a visual interpretation of each unit of verification (point), applying the forest and deforestation definitions adopted by the SMByC. This procedure is applied using Acatama tool [14] that allows defining a fixed spatial reference scale for interpretation, using all the satellite data available.

#### *Forest Degradation Around the World*

During this procedure, we also perform the classification and identify the nonclassified samples (**Figure 3**).

We used the satellite data as reference data for the sampling point interpretation for periods before and after annual composite generated in Step 6. Also, we used Google Earth Engine, Bing images, or other high-resolution image repositories if available.

## *2.2.3.4 Step 13: error matrix and confidence intervals*

Thematic accuracy assessment of the forest cover and deforestation data for the reference year is done by constructing a confusion matrix [20], using the data generated in the previous step. Subsequently, from the error matrix, a new matrix is constructed and is expressed in terms of the proportion of the estimated area.

### *2.2.4 Phase 4: calculations and reports*

To calculate the deforested area between two analysis periods, only the areas with available data in the two analysis periods are considered, so the associated (un)certainty that the event occurred in the period is analyzed.

Forest losses detected after one or several dates without information were not included in the reports in order to avoid overestimated rates due to different factors (e.g., high cloudiness or sensor failures). After each deforestation monitoring period, an analysis of consistency of the time series is performed, verifying that each pixel marked as deforestation has not been marked in the previous periods (at least 6 years) as deforested. If this is the case, the most recent result is corrected and marked as no forest (NB) or the specific area is reviewed retrospectively. The same procedure is applied for "forest recovering," maintaining the same check process in which a pixel marked as deforested could not be assigned as "forest" class until after 6 years.

**43**

**Figure 4.**

information is available at [28].

**3. Results and discussion**

**3.1 Forest cover in Colombia**

*Colombian Forest Monitoring System: Assessing Deforestation in an Environmental Complex…*

The implementation of this methodology allowed to identify the changes in forest cover (expressed in hectares) for the last 27 years in Colombia, generating reports for the years 1990, 2000, 2005, 2010, and 2012, and five annual reports for the years 2013–2017. **Figure 4** presents the workflow implemented. Additional

*Methodological workflow applied in the digital image processing protocol V.2 (Source [11]).*

In 2017, the natural forest area in Colombia was 59'312.369 ha, which represented 51.9% of the continental and insular Colombian territory (**Figure 5**). At a regional level, some provinces show high forest cover like Amazonas (97.3%), Vaupés (96.5%), and Guainía (92.9%). Likewise, other departments like Atlántico (1.4%), Sucre (2.6%), and Cesar (8.7%) have the smallest area of their territory with natural forests.

*DOI: http://dx.doi.org/10.5772/intechopen.86143*

**Figure 3.** *Acatama QGIS© tool. Verification window for sampling point interpretation.*

*Colombian Forest Monitoring System: Assessing Deforestation in an Environmental Complex… DOI: http://dx.doi.org/10.5772/intechopen.86143*

**Figure 4.** *Methodological workflow applied in the digital image processing protocol V.2 (Source [11]).*

The implementation of this methodology allowed to identify the changes in forest cover (expressed in hectares) for the last 27 years in Colombia, generating reports for the years 1990, 2000, 2005, 2010, and 2012, and five annual reports for the years 2013–2017. **Figure 4** presents the workflow implemented. Additional information is available at [28].
