Detection of Forest Tree Losses in C ô te d’Ivoire Using Drone Aerial Images

: The ﬁght against deforestation and forest degradation is now a major challenge for the preservation of global forest ecosystems. The remote sensing forest monitoring methods that are currently deployed are not always adapted to the Ivorian context because of the high cloud cover, diversity of shaded crops, and land clearing techniques. This study proposes a drone-based approach to assess intra-annual tree losses in the Boss é mati é classiﬁed forest. The method used is based on a detection analysis of tree losses in forest areas from a time series of aerial images acquired by drones from November 2018 to April 2019 on ﬁve sites in the studied forest. Based on photogrammetric models and photointerpretation, tree heights and tree crown sizes were estimated. Then, tree losses were detected based on the variation of tree heights during the study period. An analysis of the distribution of tree heights in Boss é mati é classiﬁed forest reveals that the maximum tree height was 65.06 m in November 2018 and 64.07 m in April 2019 with an average tree height of 34.29–37.00 m in November 2018 and 34.63–36.88 m in April 2019. The average tree crown area, meanwhile, was estimated to be 152 m 2 . With an estimation accuracy of about 97%, these tree structural data indicate a minimum loss of 107 trees corresponding to a clearing area of 2 ha across all the surveyed sites from November 2018 to April 2019. This forest monitoring approach shows a considerable local loss of biodiversity and should be involved in the implementation of preservation, rehabilitation, and deployment strategies in an operational deforestation monitoring system in C ô te d’Ivoire.


Introduction
Tropical forests are the focus of important global issues related to the preservation of biodiversity, climate change, and sustainable development [1]. In West Africa, the conversion of forest areas into farmland over the 2000-2010 period has been estimated at 19% by the Food and Agriculture Organization (FAO) and represents a loss of about 870,000 ha per year [2]. In Côte d'Ivoire, more than 40% of the forests disappeared within 25 years (between 1990 and 2015), with forest cover decreasing from 7.8 million ha in 1990 to 5.1 million ha in 2000 and then to 3.4 million ha in 2015, representing about 11% of the national territory [3]. To face this challenge, Côte d'Ivoire has been involved in the international mechanism for reducing emissions from deforestation and forest degradation, sustainable management of natural resources, enhancement of forest carbon stocks, and conservation of forests (REDD+) since June 2011. A new strategy for forest preservation, belt-named due to the high cocoa production here in the 1980s [12]. It is a classified forest of Côte d'Ivoire that is still well-preserved despite being subject to very strong pressures [12]. Moreover, it hosts the last population of forest elephants (Loxodonta africana) of the southeast. The vegetation is that of the Guinean domain characterized by dense evergreen rainforest. Dense forests represented 83% of the classified forest in 2016, while degraded forests represented 8%, and perennial crops of cocoa, Theobroma cacao (Sterculiaceae); coffee, Coffea arabusca and Coffea canephora (Rubiaceae); and rubber, Hevea Brasiliensis (Euphorbiaceae), represented 9% at the same date [12]. Within this forest, lianas are less numerous; the herbaceous stratum is generally well represented, with a predominance of Poaceae and Acanthaceae. The vegetation is of the semideciduous type as defined by [30] with rare epiphytes. This vegetation can be described as Malvaceae and Cannabaceae [31]. It is populated in abundance by Celtis mildbraedii (Cannabaceae), Nesogordonia papaverifera (Malvaceae), Triplochiton scleroxylon (Malvaceae), and Mansonia altissima (Malvaceae).

Overall Methodology
For the implementation of this study, a sampling protocol was set up that combines photogrammetry, stereoscopy, and change analysis techniques using very high spatial resolution images acquired with a drone (Figure 2).

Overall Methodology
For the implementation of this study, a sampling protocol was set up that combines photogrammetry, stereoscopy, and change analysis techniques using very high spatial resolution images acquired with a drone (Figure 2).

Sampling Plan
Eleven test sites were identified with the assistance of ecoguards to conduct this study in the Bossématié classified forest (Figure 1). The flights were conducted during the dry season at these sites during three acquisition campaigns: in November 2018 and April 2019. These sites are characterized by their 2016 land-use type, allowing us to evaluate the area's level of deforestation and degradation, as well as its accessibility (Table 1). The level of deforestation and degradation of these sites varies from 2% to 41% [12].

Sampling Plan
Eleven test sites were identified with the assistance of ecoguards to conduct this study in the Bossématié classified forest (Figure 1). The flights were conducted during the dry season at these sites during three acquisition campaigns: in November 2018 and April 2019. These sites are characterized by their 2016 land-use type, allowing us to evaluate the area's level of deforestation and degradation, as well as its accessibility (Table 1). The level of deforestation and degradation of these sites varies from 2% to 41% [12]. In this study, a DJI Mavic Pro drone ( Figure 3) was used. It is a multirotor system with four motors (quadcopter) powered by a LiPo (lithium polymer) smart battery with a capacity of 3830 mAh and an autonomy of 27 min. It features a wingspan of 20 cm, a weight of 736 g, and a maximum speed of 65 km/h. The DJI Mavic Pro is equipped with a 12-megapixel 4 K camera (stabilized by a pod). This camera acquires images in true colors with 8-bit radiometric resolution. This acquisition is carried out automatically using a cadence previously defined during the flight preparation. The system uses autonomous ultrasonic sensor flight technology to reduce the risk of accidents and is equipped with a GPS-GLONASS location system. The system includes a ground control radio station (connected to a smartphone) with a range of 7.3 km under normal conditions (no obstacle to the transmission of the radio signal) and a battery life of 1.5 h. In this study, a DJI Mavic Pro drone ( Figure 3) was used. It is a multirotor system with four motors (quadcopter) powered by a LiPo (lithium polymer) smart battery with a capacity of 3830 mAh and an autonomy of 27 min. It features a wingspan of 20 cm, a weight of 736 g, and a maximum speed of 65 km/h. The DJI Mavic Pro is equipped with a 12-megapixel 4 K camera (stabilized by a pod). This camera acquires images in true colors with 8-bit radiometric resolution. This acquisition is carried out automatically using a cadence previously defined during the flight preparation. The system uses autonomous ultrasonic sensor flight technology to reduce the risk of accidents and is equipped with a GPS-GLONASS location system. The system includes a ground control radio station (connected to a smartphone) with a range of 7.3 km under normal conditions (no obstacle to the transmission of the radio signal) and a battery life of 1.5 h.
The configuration of the drone and the flight planning were conducted with a DJI Go 4 (version 4.3.28) and DroneDeploy (version 4.0.0) software, respectively.

Acquisition of Aerial Images by Drone
Aerial images were acquired at a flying altitude of 200 m with a ground resolution of 6 cm/pixel and a footprint of 240 × 180 m per image. These images were acquired in the visible light spectrum (red, green, and blue) and have already been georeferenced (the geographical location of the center of each image is known). A rate of one image every 42 m allowed us to obtain a longitudinal coverage of 65%. A spacing of 95 m between the flight lines allowed for a lateral overlap of 75%. These significant overlaps ensure proper image stitching in the production of an orthomosaic according to photogrammetric and stereoscopic principles [28,29]. The entire set of 11 flight plans, each constituting nine parallel flight lines or transects in a north-south direction, allowed us to fly over an area of between 130 and 132 ha ( Figure 4). The configuration of the drone and the flight planning were conducted with a DJI Go 4 (version 4.3.28) and DroneDeploy (version 4.0.0) software, respectively.

Acquisition of Aerial Images by Drone
Aerial images were acquired at a flying altitude of 200 m with a ground resolution of 6 cm/pixel and a footprint of 240 × 180 m per image. These images were acquired in the visible light spectrum (red, green, and blue) and have already been georeferenced (the geographical location of the center of each image is known). A rate of one image every 42 m allowed us to obtain a longitudinal coverage of 65%. A spacing of 95 m between the flight lines allowed for a lateral overlap of 75%. These significant overlaps ensure proper image stitching in the production of an orthomosaic according to photogrammetric and stereoscopic principles [28,29]. The entire set of 11 flight plans, each constituting nine parallel flight lines or transects in a north-south direction, allowed us to fly over an area of between 130 and 132 ha ( Figure 4).

Orthomosaic and Digital Surface Model Production
Our methodology for processing and analyzing the aerial images ( Figure 2) was based on the principles of photogrammetry and stereoscopy. The first step was performed in the Agisoft Photoscan Professional software (version 1.4.0) and includes orthorectification and mosaicking steps. The process is subdivided into the alignment of aerial images, the production of a scatter plot, gridding, and texturing, allowing us to generate a digital surface model (DSM) and an orthomosaic [24,28,32]. As the images are located by GPS, alignment aims to reconstruct the acquisition geometry of the aerial images from the identification of link points; point cloud aims to produce a 3D model, and mesh aims to reconstruct an analysis grid to facilitate orthomosaic production; finally, texturization aims to create the texture for the 3D model [33]. The orthomosaics and DSMs were cleaned to remove poor (blurred) quality due to edge effects. The quality of photogrammetric processing was evaluated following several parameters, including the percentage of aerial image alignment and projection errors associated with the aerial image stitching process.

Geometric Corrections
In a change analysis approach, it is necessary to considerably reduce geometric errors. Geometric errors are related to uncontrolled movements (e.g., wind force) of the drone with respect to the previously defined flight line and to variations in altitude during shooting [34]. The digital surface models obtained were therefore georeferenced. The method used is image-to-image georeferencing: the DSM at date T2 is georeferenced from the DSM at date T1 on all the sites studied. The transformation applied was polynomial, and the resampling method was nearest neighbor.

Study Site Portion Delineation and Extraction
The orthomosaics and DSMs produced generally have some artifacts and blurred pixels, mostly along image boundaries. To guarantee the good quality of the image intended for processing, it is necessary to delimit and extract portions of the different study sites; this is to get rid of the poor quality (blurring) due to edge effects ( Figure 5). Thus, from an initial area varying between 130 and 132 ha for each site, the areas retained at this step for Sites 1, 6, 8, 9, and 11 are 87, 94, 97, 102, and 121 ha, respectively.

Orthomosaic and Digital Surface Model Production
Our methodology for processing and analyzing the aerial images ( Figure 2) was based on the principles of photogrammetry and stereoscopy. The first step was performed in the Agisoft Photoscan Professional software (version 1.4.0) and includes orthorectification and mosaicking steps. The process is subdivided into the alignment of aerial images, the production of a scatter plot, gridding, and texturing, allowing us to generate a digital surface model (DSM) and an orthomosaic [24,28,32]. As the images are located by GPS, alignment aims to reconstruct the acquisition geometry of the aerial images from the identification of link points; point cloud aims to produce a 3D model, and mesh aims to reconstruct an analysis grid to facilitate orthomosaic production; finally, texturization aims to create the texture for the 3D model [33]. The orthomosaics and DSMs were cleaned to remove poor (blurred) quality due to edge effects. The quality of photogrammetric processing was evaluated following several parameters, including the percentage of aerial image alignment and projection errors associated with the aerial image stitching process.

Geometric Corrections
In a change analysis approach, it is necessary to considerably reduce geometric errors. Geometric errors are related to uncontrolled movements (e.g., wind force) of the drone with respect to the previously defined flight line and to variations in altitude during shooting [34]. The digital surface models obtained were therefore georeferenced. The method used is image-to-image georeferencing: the DSM at date T2 is georeferenced from the DSM at date T1 on all the sites studied. The transformation applied was polynomial, and the resampling method was nearest neighbor.

Study Site Portion Delineation and Extraction
The orthomosaics and DSMs produced generally have some artifacts and blurred pixels, mostly along image boundaries. To guarantee the good quality of the image intended for processing, it is necessary to delimit and extract portions of the different study sites; this is to get rid of the poor quality (blurring) due to edge effects ( Figure 5). Thus, from an initial area varying between 130 and 132 ha for each site, the areas retained at this step for Sites 1, 6, 8, 9, and 11 are 87, 94, 97, 102, and 121 ha, respectively.

Digital Terrain Model Production
In this study, the DSM was used to produce the digital terrain model (DTM) following three steps. The first step consisted of partially eliminating the trees (pseudo DTM) by resampling the DSM at 10 m resolution while keeping the minimum elevation values (ground elevation). The second step consisted of a circular convolutional filter of minimum type whose size 13 was chosen after various tests. This filter size corresponds exactly to 6 pixels on either side of the central pixel (i.e., a filter with a radius of 60 m). This second step allows us to eliminate all the trees while keeping the ground elevations. At this stage, the result is a DTM with a spatial resolution of 10 m. Finally, the last step consisted of resampling the generated DTM at a 20 cm resolution.

Generation of Digital Tree Height and Canopy Models
In the field of forestry, the distribution of vegetation in general is known as digital height model (DHM). The distribution of tree heights is also known as digital canopy model (DCM). The DHM was calculated for each of the priority monitoring sites in the Bossématié classified forest based on two other landscape indicators: the DSM and the DTM according to Equation (1).
The DTM produced may have heights that do not always correspond to tree heights but rather to low vegetation heights (grass or cocoa crop) ( Figure 6). Therefore, the DCM was calculated from the DHM based on the threshold value of minimum forest tree height excluding perennial cocoa-coffee crops. This threshold value will be defined from the statistical analysis of the vegetation height distribution: where ρ is the height threshold value obtained with the drone to ensure that only forest trees are detected.

Digital Terrain Model Production
In this study, the DSM was used to produce the digital terrain model (DTM) following three steps. The first step consisted of partially eliminating the trees (pseudo DTM) by resampling the DSM at 10 m resolution while keeping the minimum elevation values (ground elevation). The second step consisted of a circular convolutional filter of minimum type whose size 13 was chosen after various tests. This filter size corresponds exactly to 6 pixels on either side of the central pixel (i.e., a filter with a radius of 60 m). This second step allows us to eliminate all the trees while keeping the ground elevations. At this stage, the result is a DTM with a spatial resolution of 10 m. Finally, the last step consisted of resampling the generated DTM at a 20 cm resolution.

Generation of Digital Tree Height and Canopy Models
In the field of forestry, the distribution of vegetation in general is known as digital height model (DHM). The distribution of tree heights is also known as digital canopy model (DCM). The DHM was calculated for each of the priority monitoring sites in the Bossématié classified forest based on two other landscape indicators: the DSM and the DTM according to Equation (1).
The DTM produced may have heights that do not always correspond to tree heights but rather to low vegetation heights (grass or cocoa crop) ( Figure 6). Therefore, the DCM was calculated from the DHM based on the threshold value of minimum forest tree height excluding perennial cocoa-coffee crops. This threshold value will be defined from the statistical analysis of the vegetation height distribution: where ρ is the height threshold value obtained with the drone to ensure that only forest trees are detected.

Digital Terrain Model Production
In this study, the DSM was used to produce the digital terrain model (DTM) following three steps. The first step consisted of partially eliminating the trees (pseudo DTM) by resampling the DSM at 10 m resolution while keeping the minimum elevation values (ground elevation). The second step consisted of a circular convolutional filter of minimum type whose size 13 was chosen after various tests. This filter size corresponds exactly to 6 pixels on either side of the central pixel (i.e., a filter with a radius of 60 m). This second step allows us to eliminate all the trees while keeping the ground elevations. At this stage, the result is a DTM with a spatial resolution of 10 m. Finally, the last step consisted of resampling the generated DTM at a 20 cm resolution.

Generation of Digital Tree Height and Canopy Models
In the field of forestry, the distribution of vegetation in general is known as digital height model (DHM). The distribution of tree heights is also known as digital canopy model (DCM). The DHM was calculated for each of the priority monitoring sites in the Bossématié classified forest based on two other landscape indicators: the DSM and the DTM according to Equation (1).
The DTM produced may have heights that do not always correspond to tree heights but rather to low vegetation heights (grass or cocoa crop) ( Figure 6). Therefore, the DCM was calculated from the DHM based on the threshold value of minimum forest tree height excluding perennial cocoa-coffee crops. This threshold value will be defined from the statistical analysis of the vegetation height distribution: where ρ is the height threshold value obtained with the drone to ensure that only forest trees are detected. Figure 6. Illustration of the principle of creating the MNH from the MNS and MNT (modified by [35]).

Estimation of Tree Crown Area
The crown is the part of the tree that is on top of the bole, including the branches, the different twigs, and the leaves [36] (Figure 7). The estimation of the ground area of the tree crown was based on photointerpretation of a sample of 124 trees based on orthomosaics from November 2018 and April 2019. First, the tree crowns were manually delineated in the QGIS 3.0 software. Then, the area (Shp) corresponding to the orthogonal projection of the tree crown to the ground was automatically calculated with QGIS 3.0 software.
Drones 2022, 6, x FOR PEER REVIEW 8 of 20 Figure 6. Illustration of the principle of creating the MNH from the MNS and MNT (modified by [35]).

Estimation of Tree Crown Area
The crown is the part of the tree that is on top of the bole, including the branches, the different twigs, and the leaves [36] (Figure 7). The estimation of the ground area of the tree crown was based on photointerpretation of a sample of 124 trees based on orthomosaics from November 2018 and April 2019. First, the tree crowns were manually delineated in the QGIS 3.0 software. Then, the area (Shp) corresponding to the orthogonal projection of the tree crown to the ground was automatically calculated with QGIS 3.0 software. Figure 7. Illustration of the principle of estimating the tree crown area (modified by [35]).

Tree Loss Detection and Validation
Tree loss detection is based on the analysis of difference in vegetation heights (∆MNH) and based on the criteria of tree height and tree crown area. The change in vegetation height is calculated using Equation (3): where t1 and t2 represent the two selected acquisition dates (November 2018 and April 2019, respectively). Negative values correspond to vegetation losses. Thus, the average value of the tree crown area allows us to discriminate between what corresponds to tree losses.
The tree height distribution maps in April 2019 and tree loss during the period of November 2018 to April 2019 were evaluated based on the photointerpretation of a sample of 511 observation points categorized as tree loss (total of 107 points), stable forest (total of 200 points), and stable nonforest (total of 204 points). Accuracy indicators, such as overall accuracy and kappa index, were finally calculated.

Quality of Photogrammetric Processing
The aerial image orthorectification and mosaicking process involved a total of 33 flights, with three diachronic flights (November 2018, January 2019, and April 2019) for each of the 11 study sites. A total of 33 orthomosaics and DSMs were produced with different levels of satisfaction. The proportions of correctly aligned images and reprojection errors from the Photoscan software ranged from 59% to 100% and from 0.551 pixels to 0.813 pixels across all study sites ( Table 2). The proportion of orthomosaics not retained in this study for reasons of poor quality (blurred images, presence of areas without data, Figure 7. Illustration of the principle of estimating the tree crown area (modified by [35]).

Tree Loss Detection and Validation
Tree loss detection is based on the analysis of difference in vegetation heights (∆MNH) and based on the criteria of tree height and tree crown area. The change in vegetation height is calculated using Equation (3): where t1 and t2 represent the two selected acquisition dates (November 2018 and April 2019, respectively). Negative values correspond to vegetation losses. Thus, the average value of the tree crown area allows us to discriminate between what corresponds to tree losses.
The tree height distribution maps in April 2019 and tree loss during the period of November 2018 to April 2019 were evaluated based on the photointerpretation of a sample of 511 observation points categorized as tree loss (total of 107 points), stable forest (total of 200 points), and stable nonforest (total of 204 points). Accuracy indicators, such as overall accuracy and kappa index, were finally calculated.

Quality of Photogrammetric Processing
The aerial image orthorectification and mosaicking process involved a total of 33 flights, with three diachronic flights (November 2018, January 2019, and April 2019) for each of the 11 study sites. A total of 33 orthomosaics and DSMs were produced with different levels of satisfaction. The proportions of correctly aligned images and reprojection errors from the Photoscan software ranged from 59% to 100% and from 0.551 pixels to 0.813 pixels across all study sites ( Table 2). The proportion of orthomosaics not retained in this study for reasons of poor quality (blurred images, presence of areas without data, artifacts, etc.) was 24% (i.e., 8 out of the 33 orthomosaics produced). Following this evaluation, Sites 1, 6, 8, 9, and 11 were selected because their orthomosaics and DSMs were of good quality and were available for both November 2018 and April 2019.

Orthomosaics, Digital Surface Models, and Digital Terrain Models
The orthomosaics obtained at each of the study sites after photogrammetric processing ( Figure 8A) were exported from the Photoscan software with a spatial resolution of 6 cm/pixel. The digital surface models ( Figure  ation, Sites 1, 6, 8, 9, and 11 were selected because their orthomosaics and DSMs were of good quality and were available for both November 2018 and April 2019.

Orthomosaics, Digital Surface Models, and Digital Terrain Models
The orthomosaics obtained at each of the study sites after photogrammetric processing ( Figure 8A) were exported from the Photoscan software with a spatial resolution of 6 cm/pixel. The digital surface models ( Figure  In Site 6, these clearings are more accentuated in the northeastern and southwestern areas. In Site 8, analysis shows that the forest is less degraded compared with other sites. Slight degradation is observed in the southeastern extremity of the site. Finally, Sites 9 and 11 are the most degraded sites with significant vegetation clearing.   To ensure the quality of the digital terrain models, we compared the DTMs generated from the two different dates selected in the study. As an example, Figure 10 shows elevation profiles that show that the DTMs are comparable from one date to another. However, there is an error in z. Since we subsequently computed the digital elevation model from the DSM and DTM, the effect of this error is significantly mitigated [37].  To ensure the quality of the digital terrain models, we compared the DTMs generated from the two different dates selected in the study. As an example, Figure 10 shows elevation profiles that show that the DTMs are comparable from one date to another. However, there is an error in z. Since we subsequently computed the digital elevation model from the DSM and DTM, the effect of this error is significantly mitigated [37].  To ensure the quality of the digital terrain models, we compared the DTMs generated from the two different dates selected in the study. As an example, Figure 10 shows elevation profiles that show that the DTMs are comparable from one date to another. However, there is an error in z. Since we subsequently computed the digital elevation model from the DSM and DTM, the effect of this error is significantly mitigated [37].

Distribution of Vegetation Heights
The

Distribution of Vegetation Heights
The

Distribution of Tree Heights
The trees here correspond to vegetation with a minimum height of 25 m. Thus, the distribution of tree heights (Table 4)

Tree Crown Area
A total of 124 tree crown trimming polygons were delineated ( Figure 13). Tree crown area values ranged from 7 m 2 to 838 m 2 , with an average of 152 m 2 across the study sites (Table 5). This average tree crown size corresponds approximately to the pixel size of a Sentinel sensor at a 10 m spatial resolution.

Distribution of Tree Heights
The trees here correspond to vegetation with a minimum height of 25 m. Thus, the distribution of tree heights (Table 4)

Tree Crown Area
A total of 124 tree crown trimming polygons were delineated ( Figure 13). Tree crown area values ranged from 7 m 2 to 838 m 2 , with an average of 152 m 2 across the study sites (Table 5). This average tree crown size corresponds approximately to the pixel size of a Sentinel sensor at a 10 m spatial resolution.

Detection of Tree Losses by Difference in Vegetation Heights
Tree loss detection is performed based on the difference in vegetation heights (Delta DHM) between November 2018 and April 2019, the minimum tree height (25 m), and the average tree crown area (152 m 2 ). Negative values represent tree losses that correspond to new clearing or tree burning ( Figure 14). The identification of tree losses is illustrated on Figure 15 Figure 13. Tree crown area delineated on aerial images acquired by drone.

Detection of Tree Losses by Difference in Vegetation Heights
Tree loss detection is performed based on the difference in vegetation heights (Delta DHM) between November 2018 and April 2019, the minimum tree height (25 m), and the average tree crown area (152 m 2 ). Negative values represent tree losses that correspond to new clearing or tree burning ( Figure 14). The identification of tree losses is illustrated on

Detection of Tree Losses by Difference in Vegetation Heights
Tree loss detection is performed based on the difference in vegetation heights (Delta DHM) between November 2018 and April 2019, the minimum tree height (25 m), and the average tree crown area (152 m 2 ). Negative values represent tree losses that correspond to new clearing or tree burning ( Figure 14). The identification of tree losses is illustrated on

Quality Assessment of Maps
The quality assessment of all the maps reveals that the overall accuracy (OA) is 97% and the kappa index is 0.95 ( Table 6). The user accuracies are estimated at 93% for tree loss, 98% for stable forest, and 99% for stable nonforest. Producer accuracies are estimated at 98% for tree loss, 97% for stable forest, and 97% for stable nonforest.

Quality Assessment of Maps
The quality assessment of all the maps reveals that the overall accuracy (OA) is 97% and the kappa index is 0.95 ( Table 6). The user accuracies are estimated at 93% for tree loss, 98% for stable forest, and 99% for stable nonforest. Producer accuracies are estimated at 98% for tree loss, 97% for stable forest, and 97% for stable nonforest.

Clearing Timing in Classified Forests and Data Acquisition Strategy
Knowledge of the clearing process in classified forest was necessary for a good spatiotemporal distribution of the drone acquisition campaigns (Figure 16). Clearing is pro-gressive and discrete and is characterized by an anthropic cocoa farming process that can be subdivided into four (4) stages.
cocoa plants with new seedlings and increase the area of their fields.
November to April: Illegal farmers set fire to the base of large trees to cause their death; this is the beginning of tree losses. During this phase which last all along the dry season, it becomes possible to detect such disturbances via optical remote sensing. The trees first lose their foliage, allowing a good amount of light to penetrate for crop growth, before falling (windfall) under the effect of the wind or being cut down with a chainsaw. The resulting cleared areas are usually considerable with a significant loss of biodiversity. It is therefore during this period that overflights must be intensified to detect the burning of the first standing trees and to alert the public to send deterrent patrols.

Quality of Photogrammetric Models and Maps
The use of the Mavic Pro drone allowed us to map the dynamics of clearings in the Bossématié classified forest on test sites following a methodology already proven in several past studies [28,32,38]. The results of the photogrammetric processing demonstrate that the selected flight altitude and image overlap levels avoid the difficulties very often encountered in the aerial image mosaicking phase, especially in forest areas [26,28]. The quality of the orthomosaics and of the selected digital surface models (reprojection errors From May to August: illegal farmers either broadcast cocoa beans into the forest understory or directly remove the forest understory before direct seeding of beans under the canopy. The first approach is becoming more and more prevalent in the Bossématié classified forest because it is almost impossible to notice the presence of cocoa crops at this stage, whereas in the second case, the ecoguards quickly identify the presence of illegal activities. From August to September: Once the seedlings appear (2 to 3 months later), they proceed to cut down the undergrowth, leaving the cocoa crop in sight. During this phase, which takes place during the short dry season, the seedlings are maintained. Until then, it is impossible to detect such a disturbance via optical remote sensing (this is the beginning of forest degradation). From September to October: As the rains pick up again, illegal farmers replace dead cocoa plants with new seedlings and increase the area of their fields.
November to April: Illegal farmers set fire to the base of large trees to cause their death; this is the beginning of tree losses. During this phase which last all along the dry season, it becomes possible to detect such disturbances via optical remote sensing. The trees first lose their foliage, allowing a good amount of light to penetrate for crop growth, before falling (windfall) under the effect of the wind or being cut down with a chainsaw. The resulting cleared areas are usually considerable with a significant loss of biodiversity. It is therefore during this period that overflights must be intensified to detect the burning of the first standing trees and to alert the public to send deterrent patrols.

Quality of Photogrammetric Models and Maps
The use of the Mavic Pro drone allowed us to map the dynamics of clearings in the Bossématié classified forest on test sites following a methodology already proven in several past studies [28,32,38]. The results of the photogrammetric processing demonstrate that the selected flight altitude and image overlap levels avoid the difficulties very often encountered in the aerial image mosaicking phase, especially in forest areas [26,28]. The quality of the orthomosaics and of the selected digital surface models (reprojection errors between 0.551 pixels and 0.813 pixels) is sufficient to assess the changes in the forest surface with good accuracy. Indeed, refs. [19,39] have shown that errors of 1 to 2 pixels are acceptable. These systematic errors in orthomosaics and digital surface models originate from the instability of the mini-UAVs and the distortion of the digital cameras used [40]. However, image overlap levels in this study (65% and 75%, frontal and lateral) could be increased to improve the quality of orthomosaics. Indeed, several studies show high levels of overlap, such as [28] (90% front overlap), [32] (80% front overlap), and [41] (90% front overlap).
In terms of the production of the set of maps of tree height distribution (April 2019) and tree loss (November 2018-April 2019) in the Bossématié classified forest, the validation reveals an overall accuracy of 97% and a kappa coefficient of 0.95. This indicates that these maps have satisfactory quality levels [42]. This shows the effectiveness of the method used in the detection of losses using aerial images acquired by drone.

Advantages and Methodological Limitations
In terms of the choice of drone, the small size of the drone used (19.9 cm × 8.3 cm × 8.3 cm) allowed us to transport and launch it from any location in the forest while avoiding trees. Compared with other acquisition platforms, such as satellites or airplanes, the drone used allowed us to acquire images at a higher resolution with a lower operating cost [43]. Although the drone can be affected by cloud cover or fog like other optical sensors, it offered us the flexibility to avoid these. However, the main limitation with the drone used is the size of the area to be covered. Indeed, although, in most cases, the forest areas to be covered easily exceed 1 km 2 (i.e., 100 ha), it should be noted that the flight capacity of a multirotor drone rarely exceeds 1 to 2 h of flight. This flight endurance being the greatest limiting factor for forestry use, we must accept that the use of a drone will remain a tool for analysis at the local scale only [16].
In terms of tree crown delineation, this method could be tedious due to manual delineation. Given the number of trees in a forest, this is probably not a viable solution for a national-scale forest preservation project.
In terms of geometric correction of the digital surface models, we propose frame-toframe georeferencing to reduce errors. The accuracy of the photogrammetric data could have been improved by using ground control points [41].

Drone versus Sentinel Sensors and Perspectives
The Sentinel-1 and 2 satellite constellations in operation provide multispectral imagery at a very high spatial resolution of 10 m. This opens new possibilities for the mapping and multiple annual monitoring of land use and land cover [12,44]. As shown in Figure 17, these individual sensors may have different potentials for monitoring forest ecosystems. Sentinel-1 radar images do not offer the possibility to distinguish clearings in the middle of a tropical forest, as the radar signal is saturated. Sentinel-2 images can be used to monitor large areas of forest clearing, but their quality is dependent on cloud cover, which is almost always present in tropical forests. Aerial images acquired by drone have strong potential for monitoring forest disturbances. The combination of these three systems should be used in the estimation of forest degradation activities in the reducing emissions from deforestation and forest degradation (REDD+) project.
The drone-based approach discussed in this study can be improved by integrating Lidar solutions and solutions related to the new generation of planet satellites. Indeed, despite the constraints related to cloud cover, the daily revisit rate of planet microsatellites allows us to obtain cloud-free mosaics with a spatial resolution of 3 m to 5 m [45]. The use of planet images would allow for the refinement of forest disturbance early detection using drone data. Lidar solutions should be tested for the early detection of forest understory loss [46]. This is a major limitation of current remote sensing tools and methods to detect cocoa plots under shade. This would allow for intervention on the ground in the early stages of illegal clearings by cocoa farmers infiltrating classified forests and protected areas. The technologies mentioned above would allow Côte d'Ivoire to have innovative tools adapted to its agroforestry context and for the implementation of its current ambitious sustainable development objectives. The drone-based approach discussed in this study can be improved by integrating Lidar solutions and solutions related to the new generation of planet satellites. Indeed, despite the constraints related to cloud cover, the daily revisit rate of planet microsatellites allows us to obtain cloud-free mosaics with a spatial resolution of 3 m to 5 m [45]. The use of planet images would allow for the refinement of forest disturbance early detection using drone data. Lidar solutions should be tested for the early detection of forest understory loss [46]. This is a major limitation of current remote sensing tools and methods to detect cocoa plots under shade. This would allow for intervention on the ground in the early stages of illegal clearings by cocoa farmers infiltrating classified forests and protected areas. The technologies mentioned above would allow Côte d'Ivoire to have innovative tools adapted to its agroforestry context and for the implementation of its current ambitious sustainable development objectives.

Conclusions
This study aimed at the early detection of tree losses in the forests of Côte d'Ivoire using aerial photographs taken by drones. The methodological approach was based on the use of aerial images acquired by drones (spatial resolution 6 to 12 cm) to assess tree losses in the classified forest of Bossématié. The choice of this classified forest is justified by the low annual deforestation rate recorded between 2016 and 2019, which is estimated at 0.35%. This approach allowed us to evaluate tree losses in this relatively well-preserved classified forest subjected to strong anthropic pressures with satisfactory accuracy. Indeed, it allowed for the detection of a minimum of 107 tree losses corresponding to a clearing area of 2 ha on all sites studied in the classified forest of Bossématié during the period from November 2018 to April 2019 with a satisfactory overall accuracy of 97%. Therefore, this study shows the interest in using drones in the management and monitoring of classified forests and protected areas in Côte d'Ivoire. The drone is therefore a precise tool whose capabilities are constantly improving. In addition to its proven value in terms of early detection of the first standing tree burns, regular drone overflights are also proving to be a strong deterrent to cocoa farmers infiltrating protected forests, who realize that the forest is under close surveillance.
The results of this work must be integrated into the development and management plans of classified forests and protected areas. Indeed, the use of this spatial tool (the drone) could contribute to the good management of these protected areas through better monitoring of the land (fine and regular mapping, detection of forest disturbances, and early warning of clearings).

Conclusions
This study aimed at the early detection of tree losses in the forests of Côte d'Ivoire using aerial photographs taken by drones. The methodological approach was based on the use of aerial images acquired by drones (spatial resolution 6 to 12 cm) to assess tree losses in the classified forest of Bossématié. The choice of this classified forest is justified by the low annual deforestation rate recorded between 2016 and 2019, which is estimated at 0.35%. This approach allowed us to evaluate tree losses in this relatively well-preserved classified forest subjected to strong anthropic pressures with satisfactory accuracy. Indeed, it allowed for the detection of a minimum of 107 tree losses corresponding to a clearing area of 2 ha on all sites studied in the classified forest of Bossématié during the period from November 2018 to April 2019 with a satisfactory overall accuracy of 97%. Therefore, this study shows the interest in using drones in the management and monitoring of classified forests and protected areas in Côte d'Ivoire. The drone is therefore a precise tool whose capabilities are constantly improving. In addition to its proven value in terms of early detection of the first standing tree burns, regular drone overflights are also proving to be a strong deterrent to cocoa farmers infiltrating protected forests, who realize that the forest is under close surveillance.
The results of this work must be integrated into the development and management plans of classified forests and protected areas. Indeed, the use of this spatial tool (the drone) could contribute to the good management of these protected areas through better monitoring of the land (fine and regular mapping, detection of forest disturbances, and early warning of clearings).