Early warning tropical forest loss alerts in Peru using Landsat

Since March 16, 2017, the National Forest Conservation Program for Climate Change Mitigation (PNCBMCC) of Peru’s Ministry of the Environment (MINAM) has been implementing a methodology to detect early warning alerts of humid tropical forest cover loss in Peru using data from the Landsat 7 and 8 satellites. The method uses Direct Spectral Unmixing (DSU) to detect forest loss as small as 25% of a pixel. Between March 16 and December 25 of 2017, 500 Landsat images have been used to detect 137,143 hectares of humid tropical forest cover loss, including deforestation for agricultural expansion and illegal or informal extractive activities, such as the opening of roads for selective logging. Natural forest loss was also detected, produced by windstorms and landslides in mountainous areas, among others. The results were verified with high-resolution satellite images and the accuracy was evaluated using a stratified random sample, showing a high level of both user’s and producer’s accuracy. The early warning alerts are distributed and available through the Geobosques platform (http://geobosques.minam.gob.pe).


Introduction
Global warming has been of increasing concern in recent years, produced by higher greenhouse gas (GHG) emissions from different economic sectors at a global scale. In Peru, the land use, land use change, and forestry sectors (LULUCF) contribute the most emissions, representing 51% of total emissions (MINAM 2016), and primarily due to the deforestation of humid Amazonian forests. This dynamic has incentivized public and private institutions to develop, propose and execute research on remote sensing for forest monitoring, given that it can cover a wide area in a reasonable time and provide verifiable information for decision making on monitoring, control, and sanctioning of deforestation.
There are currently many methods for early detection and monitoring of forest loss or change using low-and medium-resolution images.  (Souza et al 2005) to detect deforestation in the state of Mato Grosso in Brazil (Souza et al 2009). Two new algorithms that use MODIS data for the rapid detection of forest disturbances (Tang et al 2019) were proposed recently. Although these algorithms have better results than Terra-i, their nationwide operation has not yet been tested. The main limitation of the use of low spatial resolution data such as MODIS (250 m) is that they are not able to detect small-scale deforestation events. In 2016, 72% (116 910 ha) of deforestation in tropical humid forests of Peru occurred in patches of less than 5 ha (http://geobosques.minam.gob.pe/geobosque/ view/perdida.php). This indicates that, MODIS would potentially detect less than 28% of the deforestation that is actually happening. In 2016, the Global Land Analysis and Discovery (GLAD) lab in the Department of Geographical Sciences at the University of Maryland presented a prototype for forest disturbance alerts in humid tropical forests for Peru, the Republic of Congo, and Kalimantan, Indonesian using data from the ETM+and OLI sensors (Hansen et al 2016). These data have a spatial resolution of 30 m, which allows detecting small-scale deforestation. Currently those alerts are available for more than 20 countries. New methodologies have reported the combined use of optical data and Synthetic Aperture Radar (SAR) in the detection of deforestation (Reiche et al 2018, Perbet et al 2019). This minimizes the bias between the date the image was taken and the date on which deforestation occurred.
In 2012, with the support of the SilvaCarbon program, the Department of Geographical Sciences at the University of Maryland began providing technical assistance to PNCMBCC for the mapping of humid tropical forest loss for the period 2000 to 2011 (Potapov et al 2014, MINAM 2015a, with support continuing in subsequent years. In December 2015, the Norwegian Agency for Development Cooperation (NORAD) approved a project for the implementation of the Joint Declaration of Intention for REDD+between Peru, Norway, and Germany, signed by the World Wildlife Fund and the PNCBMCC (DCI-WWF). Within the framework of this agreement, a preliminary method was developed for the early detection and quantification of humid Amazonian forest loss in Peru using data from Landsat 8 (Vargas et al 2017), in order to facilitate the creation of a nationally-owned forest monitoring system in the future. In parallel, in January 2016 the PNCBMCC and the World Resources Institute (WRI) signed a data sharing agreement, through which PNCBMCC received information for the GLAD alerts. In the second week of March 2017, the GLAD lab stopped producing the alerts temporarily, and the PNCBMCC had no information to report to its users, Given this lack of information, the PNCBMCC implemented its own methodology, which is an update of the preliminary method proposed by Vargas et al (2017). With this, the PNCBMCC has achieved independence in the generation of data and managed to improve the detail of detection of forest cover loss, contributing to the sustainability of monitoring of humid tropical forest in Peru.
This document describes the methodological process and results of the method used for the detection of early warning humid tropical forest alerts in Peru implemented by PNCBMCC.

Study area
The study area is Peru's humid tropical forests (figure 1), which cover an area of 78, 308, 801 hectares and represent 60.9% of the country (MINAM 2012). Peru is considered a country with a large extension of forests and a moderate annual deforestation rate, in which the forest conversion predominates mainly due to agricultural activity, artisanal mining, and industrial agriculture (Potapov et
For the period 2013-2016, two sets of the best images per year were downloaded and used to extract forest endmembers and forest loss, and for 2017 all available images between March 16 and December 25 were downloaded and used to detect the loss of forest. All images were calibrated to Top of the Atmosphere (TOA) reflectance. For the classification of clouds, haze, and shadows, a set of binary rules implemented in a decision tree was used. These rules were generated based on the spectral response of clouds, haze and shadows of 30 Landsat ETM+and OLI images. The edges of the clouds are difficult to classify, due to the spectral mixture they have with other materials. To ensure their detection, a dilation filter was applied with a 5×5 window. To standardize the projection and maintain coherence with maps of primary forest and annual forest loss generated by MINAM. All work was done in UTM, zone 18 S, datum WGS84.

Direct spectral unmixing (DSU)
To detect the early warning alerts, we developed a new method, called Direct Spectral Unmixing (DSU). DSU is based on Linear Spectral Mixing Model (LSMM), which provides quantitative information of the materials that make up the pixel. LSMM assumes that the spectral response of a pixel is the linear combination of the materials that are inside the pixel (Endmembers) ( DSU only uses endmembers of forest and forest loss, and assumes that when a pixel loses forest cover from natural or anthropogenic causes, the result may be a pixel with bare soil, a mixed pixel with soil and dry vegetation, or residuals of deforestation like tree trunks which may also be mixed in with standing forest. The endmembers used in this study were obtained from Landsat images. The advantage of extracting endmembers directly from the images is the ease with which they can be obtained, and the similarity in scale to the data (Roberts et al 1998). The forest endmember is the average of the spectral response from primary forest, obtained by taking the minimum and maximum reflectance values from the composite of images for the period 2013-2016 through a random sample of pixels that coincides with the primary forest layer of the year 2016. The endmember of forest loss is the average of the spectral signature of forest loss, obtained using a random sample of pixels with the highest probability of being deforested during the period 2013-2016. These endmembers were used to create a LSMM, which models the spectral behavior of distinct percentages of forest cover loss within a pixel (see figure 2).
A common problem of spectral unmixing is to be able to estimate the percentage of each class within a pixel (Schowengerdt 2007). DSU solves this problem in a practical way by applying the ratio between the SWIR1 and NIR bands for each percent of forest cover loss in the LSMM, obtaining thresholds that relate directly with the percent of forest cover loss within a pixel. Figure 3 shows the RapidEye image taken on 10/05/2017 with two deforested areas. In this image the percentage of loss of forest cover was obtained by applying DSU in a Landsat 8 image taken on 10/10/2017. It can be seen that, at the edges of the deforested area, DSU allowed the calculation of the percentage of loss of forest cover within the Landsat pixel.

Detection of early warning alerts
Within the forest area defined by the 2016 primary forest layer, early warning alerts were detected weekly starting at March 16, 2017. In order to achieve this, we integrated the TOA reflectance of all Landsat ETM+and OLI images into mosaics and applied a binary decision tree that includes cloud, haze and shadow classification with the detection thresholds of forest cover loss. With the OLI images, the threshold was set to detect up to 25% forest cover loss over low slopes and up to 35% forest cover loss in areas with mountainous relief. This differentiation avoids the possible detection of false positives in areas of forests with mountain shadows. The 35% threshold was also used with ETM+images. This threshold was used to avoid false positive detection due to the possible presence of low-density clouds that could not be classified in the data; these clouds are detected using the ultra-blue band, which is absent in the ETM+sensor. Finally, in the forest loss layer we deleted the pixels that intercept with clouds or haze with dilation filter with a 5×5 window.
The date on which the loss of forest cover is detected corresponds to the date of the ETM+ or OLI image used, and ideally, the data show no presence of clouds. The actual date on which the loss of forest cover occurred should not be more than 8 days prior to the date of the ETM+or OLI image used for its detection.
The early warning alerts detected each week are used as an auxiliary layer within the binary decision tree in order to avoid repeated detections in following weeks. The early warning alerts are published and distributed through the Geobosques platform of PNCBMCC. Figure 4 shows the flow diagram for the detection and distribution of early warning alerts.

Comparison with GLAD data
As mentioned in the introduction, this methodology is developed to support the sustainability of monitoring and improving the detail of detection of forest cover loss. In order to know if we were able to improve the detection of the loss of forest cover, we compared our results with the early warning alerts of the Global Land Analysis and Discovery (GLAD) lab of the UMD. These alerts reach up to 50% detection of forest cover loss within a pixel of Landsat (Hansen et al 2016). The data of the year 2017 were downloaded from the GLAD website (http://glad-forest-alert.appspot.com/) On May 15, 2018, the data were reprojected to UTM, zone 18 S, datum WGS84 and only the early warning alerts that coincided with the primary forest layer for 2016 were used.

Verification and accuracy assessment
Satellite images available on the Planet Platform were used for the verification and accuracy assessment of the early warning alerts. The costs associated with the use of this platform were covered by the DCI -WWF project.. This platform is characterized by having satellite data with high spatial and temporal resolution (Doves and RapidEye images with a spatial resolution of 3 m and 5 m respectively). Sentinel-2 10 m spatial resolution bands were also used. The use of these images ensures that the reference data used to evaluate our results is better than that applied to generate the information (Olofsson et al 2014). In addition, the use of high spatial resolution images minimizes uncertainty in the interpretation of sample units.
The early warning alerts were visually verified using available images from the Planet platform ((Dove and RapidEye). The largest deforestation events are described in official reports, where images from Planet and Sentinel-2 are used to show the before and after of a deforestation event. These reports can be found in the ATD category-monitoring files of the downloads section of the Geobosques platform (http://geobosques.minam. gob.pe/geobosque/view/descargas.php).
To understand the accuracy of forest loss detection, we used a stratified random sample. We decided to use this sampling method due to the need to know the accuracy of the early warning alerts detected in low slopes and mountainous areas, added to the advantage of having access to images of the Planet platform for the entire humid tropical forest. We created the following four strata: forest cover loss/non-forest cover loss in mountainous areas and forest cover loss/non-forest cover loss in low slopes). The stratum of forest cover loss corresponds to the early warning alerts detected until 10/18/2017, which amount to 117 240 ha of the total detected. This is because our access to the Planet Platform was available until 12/31/2017 and because the months of October, November and December correspond to the humid season, where satellite images have a high presence of clouds. By doing this, we reduced the probability of having sample units that cannot be interpreted due to the presence of clouds. The non-forest cover loss stratum corresponds to the 5-pixel buffer around the detected early warnings that coincide with the primary forest layer of 2016. This prevents us from having random sample units in non-forest areas or deforested areas before 2017, so that sampling in this stratum is restricted to areas with the highest probability of having deforestation events, which commonly occur in areas adjacent to deforested lands. Figure 5 shows an example of the forest cover loss and non-forest cover loss strata, and it also depicts how non-forest areas were excluded from the non-forest cover loss stratum. The forest cover loss and non-forest cover loss strata were subdivided using a layer of lowland forest and mountain forest that was created using the vegetation map of MINAM (MINAM 2012). In total, a single sample of 1384 sample units was used at the pixel level. The distribution of sample units was proportional to the area of pixels of the strata within the low sloped forest and forest in mountainous zones, resulting in 988 and 396 sample units, respectively. Then, these sample units were distributed equally among each stratum. Table 1 shows the number of pixels, their equivalent area and the sample units used for each stratum.
The interpretation of sample units was performed by an expert not involved in the alert generation process. Each sample unit (pixel) was interpreted using Planet (Dove and RapidEye) or Sentinel-2 images. Sample units with presence of forest cover loss were labeled with the number 1 and sample units with presence of forest were labeled with the number 0. A confusion matrix was created based on these interpreted sample units, with accuracy figures for each stratum and 95% confidence intervals.

Early warning alerts
Between March 16 and December 25 of 2017, a total of 137,143 hectares of humid tropical forest cover loss were detected. Of detected alerts, 78.8% occurred in low slope areas, while 21.2% occurred in areas with mountainous relief.
In order to know the areas with the highest density of early warning alerts, we developed a point density map -while previously, early warning alerts used to be converted to vector format-and we used a radius of 5000 m, which shows areas with the highest density of early warning alerts in red and the areas with lower density in green. The highest density of early warning alerts occurred in the regions of Madre de Dios, Ucayali, and Huánuco (see figure 6).
In Madre de Dios, the biggest patch of deforestation detected was more than 480 ha. Using visual interpretation, it was estimated that 7,713 ha of forest loss was due to illegal mining, representing 5.6% of all detected forest loss. In the case of the Ucayali and Huánuco regions, the highest density of alerts is found in the  area near the Jorge Basadre Highway and the road connecting Codo del Pozuzu and Puerto Inca. In these regions, 1360 patches of more than 5 ha were detected, attributable to large-and medium-scale agriculture and livestock activity. There were also 109 patches of deforestation larger than 20 ha, totaling 4085 ha. Based on the historical dynamics of deforestation, these areas are attributed to extensive cattle ranching and agro-industrial cultivation (MINAM 2016). Based on a weekly basis analysis, 65% of patches were smaller than 1 ha. However, if we look at patch size considering the entire study period, that figure decreases to 46%, with more patches topping 1 and even 5 ha (see figure 7). This shows the evolution of patches over time, as primary forest loss starts small and becomes larger over time.
From the above, it can be determined that there is a 19% decrease in the patches of forest cover loss smaller than 1 ha, which have been redistributed to patches larger than 1 ha, and that forest cover loss patches greater than 5 ha increased by more than 50%. For example, of the 480 ha of forest loss detected in Madre de Dios, 36 ha were detected in March, the highest deforestation happened in August and September with more than 280 ha, and deforestation continued until December (see figure 6). This indicates that the deforestation driver remains active during the humid and dry seasons.
The threshold of detection of up to 25% of forest cover loss in a pixel permits the detection of some forest degradation events such as selective logging and opening of roads. In total, 1416 km of roads wider than 7.5 meters were detected by the alerts, most of which are located in the regions of Madre de Dios, Ucayali and Loreto. Figure 8 shows examples of road detection and some selective logging events.
The highest detection of early warning alerts occurred in the dry months. This is because deforestation associated with activities such as agriculture, livestock and selective logging takes place in this season since it is easier to access a land, burn it, clear it and sow it. This also coincides with the greater availability of images with lower percentage of clouds. However, deforestation caused by illegal mining can start in the humid season.

Comparison with GLAD data
Between March 16 and December 25, 2017, it was possible to detect 137,143 ha of forest cover loss, while for the same period, GLAD detected 99,958 ha. The multitemporal comparison of the number of hectares detected by both methodologies shows in most cases the early warnings of Geobosques detected more hectares of forest cover loss, starting in the month of August (see figure 9).
GLAD detected 983 km of roads. The early warning alerts of Geobosques detected 433 km of roads that were not detected by GLAD. This shows a significant advantage in detecting the opening of roads, which are commonly used for selective logging. Figure 10 shows the detection of roads made by the early warning alerts of Geobosques and GLAD. The detection of up to 25% of forest cover loss within a pixel allowed detecting more forest cover loss by the opening of roads.
Both methods have a low detection of forest loss in the humid season and a greater detection in the dry season, since in this season there is a greater probability of obtaining images with low percentage of clouds. Table 2 shows the confusion matrix, accuracy and confidence intervals obtained for the early warning alerts. The user's and producer's accuracies for the forest cover loss stratum in mountainous areas were 97.0% and 95.0%, respectively, while for the forest cover stratum in low-slope areas the accuracies were 94.9% and 91.6%, respectively. The user accuracy in both strata is high, which means that our commission error is low, while the  producer's accuracy, which indicates the omission of forest loss pixels, has a lower value in the forest cover loss stratum in low-slope areas. This is likely due to the dilation filter that was applied to clouds, which partially or totally eliminated the forest loss found next to clouds. The overall accuracy obtained was 93.9%.

Accuracy assessment
The early warning alerts are updated weekly and distributed through the Geobosques platform from PNCBMCC of MINAM (http://geobosques.minam.gob.pe). Geobosques has more than 2500 subscribers who receive information about alerts in their areas of interest via email (see appendix).

Conclusions
The method developed provides weekly information on the loss of humid tropical forest cover in Peru. The methodology uses a large amount of Landsat data and the method developed allows for high accuracy. This assures the reliability of the data to support authorities responsible for forests with accurate information on primary forest loss.
The direct spectral unmixing (DSU) method permits the selection of a threshold for the minimum percent of forest cover loss within a pixel. This provides consistency and comparability between the data and the use of the  detection threshold of up to 25% of forest cover loss in a pixel, which allowed us to detect roads that were not detected by GLAD's early warning alerts. The analysis of forest cover loss patch size shows that 19% of the forest loss that was initially detected as less than 1 ha had become a larger area of deforestation by the end of the year. This dynamic provides an opportunity for the responsible authorities to evaluate the legality of the deforestation and take corrective action to stop deforestation in areas that don't have an authorization for land use change.
Most alerts were detected in months with the least cloud cover. To help overcome the limited availability of optical data when there is presence of clouds, the Japan International Cooperation Agency (JICA) has been working with the PNCBMCC and the National Forest and Wildlife Service (SERFOR ) to apply Synthetic Aperture Radar (SAR) data in the early detection of forest cover loss. UAS systems can also be used for monitoring small areas and support in covering gaps of information related to the presence of clouds in the optical data. The methodological process described here only uses spectral bands that are also available in Sentinel-2 data, with the thought that in the future these data can also be used in the next generation of early warning alerts.
forest cover loss with the previous year, as well as the official data on the annual loss of humid tropical forests of Peru. The viewer is only available in Spanish.
In order to promote the use of these deforestation monitoring tools in Peru, the PNCBMCC has been carrying out capacity building activities at the level of subnational governments (regional and local governments) and their allies in forest conservation (public and private institutions, indigenous organizations, among others), to improve forest management. Until July 2019, Geobosques had 1530 registered users who receive information on early warning alerts via email and are able to download these data in order to plan verification, prevention and/or control actions for activities that cause deforestation. Figure A2 shows the distribution of Geobosque users.