Dynamics of Land Use and Land Cover Change in the South Talihya Watershed North Kivu, Eastern Democratic Republic of Congo

The anthropization of forest ecosystems has become a major environmental problem with negative impacts on biodiversity around the world. Responding to such a problem requires monitoring land use. Thus, this study aims to analyze the Spatio-temporal dynamics of land use (LULC) to implement effective management strategies for forest ecosystems in the South Talihya watershed in eastern DR Congo. For this, a diachronic study of Landsat TM+ satellite images from 1987 and 2001 and Sentinel-2 from 2020 was used. The results revealed that natural formations have regressed in favor of croplands and fallow mosaics in the Talihya South watershed. The forests which occupied 183.68 km 2 in 1987, decreased to 140.94 km 2 in 2001 and 108.2 km 2 in 2020. Bare lands and buildings represented 87.54 km 2 in 2020 against 115.33 km 2 in 2001 and 58 .63 km 2 in 1987. The current state of land use indicates that Crop lands, fallows, and pastures occupy a high surface area compared to forests and bare lands and buildings. These Crop lands, fallows, and pastures occupied 317.88 km 2 in 1987, 303.74 km 2 in 2001, and 373.34 km 2 in 2020. The observed changes result from overexploitation of resources characterized by the expansion of agricultural lands and tree logging. The annual deforestation rate calculated from forested areas of Talihya South watershed between dates t 0 (1987) and t 1 (2020) was 1.60%. In view of these results, participatory land use planning is a way to ensure sustainable management of the existing residual forests in this watershed.


INTRODUCTION
The forest cover of the Democratic Republic of Congo (DR Congo) is over 166 million hectares and alone accounts for more than 62% of the Congo basin's forest cover, which is 69% dense forest as opposed to approximately 30% only of other forest types (Tchatchou et al., 2015). These forested formations in DR Congo cover about 54.6% of its surface area (Nsangua et al., 2018). These forests are also at risk of degradation not only due to illegal tree logging, but mainly following slash-and-burn cultivation for food production, inducing an annual loss of 532,000 of forest cover (Hatakiwe, 2016).
These dense tropical forests of the Congo Basin host tremendous biodiversity, constitute a source of food and livelihood for more than 60 million people and about 44 million hectares are under forest exploitation regimes as forest concessions (Megevand et al., 2013;Manguiengha et al., 2019). The primary causes of deforestation include but not limited to slash-and-burn agriculture, timber logging, the increasingneed for fuelwood, and wildfires (Hami et al., 2007;Megevand et al., 2013;Sylla et al., 2019). These factors of deforestation and forest degradation result in changes in vegetation structure and floristic composition of the landscape (Sylla et al., 2019). According to Gebhardt et al. (2015), the intensification of activities over time has generated significant remodeling, not only of the forest cover but also of the underlying soil.
The drivers of deforestation in Lubero territory are similar to those reported in the tropical zone in general. This eco-region is characterized by the scarcity of forested areas, as a direct consequence of deforestation since the 18 th century through the field and pastures establishment (Vyakuno, 2006). The watershed drained by the South Talihya River is situated in the cool highlands of Lubero, densely populated with a decreasing forest cover in the market gardening area extending from Kyondo to Masereka. The remaining forests in this zone are in form of forest refuges or forest fragments (Vyakuno, 2006), indicating anthropization in these areas. Some areas that have been deforested for centuries are struggling to evolve towards the mountain forest climax and remain at the savanna stage because of the climate and repetitive bushfires that slow down spontaneous reforestation (Vyakuno, 2006). In this context of demographic pressure, it is necessary to know the trends of the dynamics of the forest cover over time as well as its underlying drivers for informed decision-making (Akakpo et al., 2017). This raises the question of sustainable resource management in this watershed, which is the epicenter of market garden production that serves the proximate cities of Butembo and Beni.
The development of remote sensing and geographic information system (GIS) techniques allows for an increasingly precise approach to assessing land use dynamics (Oloukoi et al., 2006;Kpedenou et al., 2016). Remote sensing and GIS, which are not only increasingly used for the study of phenomena occurring on the earth's surface, form essential tools in interactive decision support systems (Diedhiou et al., 2020). This study examines the dynamics of land use in the South Talihya watershed. The mapping of land use and land cover changes are real tools for planning and decision-making support, especially in the management and preservation of natural resources and ecosystems.

Study Site
The study was conducted in the South Talihya River watershed. This watershed is located in the Cool Highlands of Lubero, in the North-East of the DR Congo. These Cool Highlands are located between 2000 and 3100 m of altitude have schistose and granitic rocks from the Precambrian period and relief of dissected and abrupt hills. The annual averages of temperature and precipitation are low due to the altitude: 15 to 17 o C and 1,110 to 1,330 mm. The forest is mountainous with orophilous species. The lower limit of this stage has as agronomic indicators potatoes and wheat that grow well only above the 1800 m isohypse (Kasay, 1988;Vyakuno, 2006). The South Talihya River originates south of Masereka and flows south before turning east to flow directly into Lake Edouard. Its basin offers the juxtaposition of different forms corresponding to different erosion cycles (Vyakuno, 2006). The South Talihya watershed is populated by the Nande, the largest ethnic group traditionally known as Yira. The economy of the region is based on subsistence agriculture that is carried out by resource-poor farmers.

Choice of satellite images, dates, and Acquisition of satellite images
To study the dynamics of land use, Landsat images were used. Since the landscape of the South Talihya watershed is highly fragmented, the 2020 Sentinel-2 image was also used to improve the accuracy of the classification.
Compared to Landsat data, the revisit time of the satellite of 16 days does not allow under certain conditions to optimize the chances of having several images in a short time for monitoring and to avoid the problem induced by the presence of clouds, recurrent in the tropical context. Following this bottleneck, the development of the ESA Copernicus program since 2015 has offered new Sentinel-2 optical images with improvements in spatial resolution (10 to 60 m) and the revisit time of the satellites (5 days). With the short revisit time, Sentinel-2 imagery allows better monitoring of vegetation and better management of cloud-induced problems (Konko et al., 2021).
Considering the specific characteristics of each satellite image and their availability, two Landsat scenes were used, one from 1987 and the other from 2001 (Enhanced Thematic Mapper+: ETM+). For 2020, the Sentinel-2 image with a resolution of 10 m was used to study the spatial dynamics of soil cover in the South Talihya watershed. The choice of dates was based on data availability and quality. In addition, we tried to coincide these dates with key dates in the development of the region. The spatial resolution of these images is 30 meters, of UTM zone 35 S projection with WGS84 reference ellipsoid. Landsat images were downloaded during the period marked by little rainfall. In fact, Hountondji (2008) and Akakpso et al. (2017) suggest that the driest months of the year are the most favorable period for vegetation characterization by spatial remote sensing. The images acquired during such periods are not very cloudy (Agbanou, 2018) and allow for to generate of bio-environmental and geographical information close to the reality of the field (Hountondji, 2008;Oszwald et al., 2012;Akakpso et al., 2017;Agbanou, 2018).
Landsat satellite images were downloaded from the Earth Explorer website (https://earthexplorer.usgs.gov/) in geoTiff format, selecting those with the least amount of disturbance, especially cloud cover. The 2020 Sentinel-2 image was downloaded from https://visioterra.org/VtWeb/. The Radar (SRTM) images, essential for the delineation of the studied watershed, were downloaded from http://dwtkns.com/srtm. The characteristics of the satellite images used are described in Table 1.

Color composition and filtering
The color composition was done depending on each image scene. The images were visualized in true colors or in false colors according to the need. An example of a color composition for the images used in this study is illustrated in Figure 1.  From the true color composition (Figure 1), the land use classes appear with natural colors: water appears in blue, woody vegetation appears in dark green, cultivated land and fallow land in light green and finally bare and built-up land in gray. In contrast, in false color (Figure 2), water is black, tree vegetation with high chlorophyll intensity is dark red, and cultivated lands and fallow lands are light red. The whitish spots visible in both images are the clouds. Once the color composition was done for each image, we performed a Gaussian filter which allowed the elimination of the isolated pixels and homogenized the thematic classification.

Extraction of the study areas
To prevent bias in the classification, areas of interest were extracted on the scene of the satellite image used. This extraction was done from the multispectral images obtained, following the boundaries of the South Talihya watershed using the Resize data tool in Basic Tools of ENVI software version 4.6.

Resampling
Resampling of the pixels was applied because the images (Landsat 5 ETM+ and Sentinel-2) did not have the same spatial resolution. This operation allows them to be superimposed during the implementation of the transition matrix. In addition to radiometric improvements, the Landsat images underwent geometric corrections that consisted of realigning the image to a geographic reference frame (Dolbec et al., 2005;Katembera Ciza et al., 2015;Tsewoue et al., 2020). The nearest neighbor resampling method was adopted in this study. This method uses the nearest pixel value without any interpolation to create the rectified pixel value. Therefore, the radiometric values of the pixels are not affected during resampling (Sagne et al., 2015).

Image classification
The supervised classification method was implemented for the processing of Landsat images. The Maximum Likelihood (ML) algorithm was chosen for image classification. This algorithm assigns each pixel to the class to which it has the highest probability of belonging (Katembera Ciza et al., 2015;Agbanou et al., 2018;Orekan et al., 2019;Manguiengha et al., 2019) by highlighting the standard error range between pixel values and those of different training sites (ROIs) (Orekan et al., 2019). The established classification as well as the representation colors of each land cover category were adapted from the Corine Land Cover (CLC) nomenclature Mbaiyetom et al., 2020). The classification of the 2020 Sentinel-2 image was done based on ground data from the same year, whereas the classification of the 1987 and 2001 Landsat images was done retrospectively and deductively (based on the spectral similarity of the 2020 Sentinel-2 image classes, due to the lack of ground data for these years). For each land cover class, ROIs (training areas) were delineated away from transition areas to avoid including mixed pixels that could be classified into two separate classes. The definition of the training areas (ROIs) on the Sentinel-2 image is shown in Figure 3 below. From the ground data, three land cover classes were identified: Forest (including forest residuals and tree plantations), Mosaic fields and fallows (including croplands and herbaceous fallows), Savannahs (including grassy formations interspersed with shrubs), and Bare lands and builtup soils (including plowed fields, roads, and builtup areas). For each occupation class, georeferenced points constituting "training areas" were used for verification. These coordinates had a spatial distribution over the entire study area as suggested by Adjonou et al. (2019). For Tsewoue et al. (2020), ground truth missions allow the validation of the boundaries of the different land use classes resulting from the interpretation and influence the accuracy of the classification. In this phase, vectorization was executed on the classified images. These were transformed into a Shapefile to determine the surface areas of each land use class.

Classification accuracy and validation
The accuracy of the classifications obtained in this paper was evaluated through the use of a confusion matrix or contingency table. For Randriamalala (2015), the confusion matrix is one of the most used statistical indicators in the performance of classification. It highlights the confrontation between the results of the supervised classification and the reference data that represent the ground truth (Soulama et al., 2015;Rifai et al., 2018;Okanga-Guay et al., 2019;Orekanet al., 2019;Gbedahi et al., 2019;Useni Sikuzani et al., 2020). In fine the model accuracy indicators (overall accuracy (OA), User accuracy (UA), Producer accuracy (PA), omission errors, and commission errors) were calculated from the confusion matrixes obtained by cross-validation. The validation of the classification was based on the Kappa index, which was assessed by means of a confusion matrix. This coefficient incorporates all the elements of the confusion matrix (Hountondji, a b c 2008) and characterizes the ratio between the number of well-ranked pixels and the total pixels sampled (Kpedenou et al., 2016;Cabala Kaleba et al., 2017;Fotso et al., 2019;Adjonou et al., 2019). It is given by the following formula (Bekele et al., 2021): With r=number of rows in the matrix; Xii= number of observations in row i and column i (diagonal elements); x+i and xi+ = represents the marginal total of rows r and columns i and N= number of observations. The Kappa coefficient, which varies from 0 to 1 (Orekan et al., 2019) and reflects a level of concordance that is all the higher the closer its value is to 1 (Hountondji, 2008). It is divided into five categories: very low agreement from 0 to 0.20; low agreement from 0.21 to 0.40; moderate agreement from 0.41 to 0.60; substantial agreement from 0.61 to 0.80; and almost perfect agreement from 0.81 to 1 (Landis and Koch, 1977). A study can be validated if the Kappa index is between 50% and 75% (Cabala Kaleba et al., 2017).

Analysis of land cover change
The classified images were transformed into a shapefile so that the areas of each land use unit could be calculated. This made it possible to establish the transition matrix as recommended by Schlaepfer (2002). It allows to highlight in a way the different forms of conversion undergone by the land use units between two dates t 0 and t 1 , and to describe the changes that occurred (Kpedenou et al., 2016;Avakoudjo et al., 2014;Arouna et al., 2016;Toko Imorou et al., 2019;Oloukoi et al., 2006;Avakoudjo et al., 2014;Ayena et al., 2017;Samba et al., 2021;Bamba et al., 2008).
The transition matrix is in the form of a square matrix and consists of X rows and Y columns (Avakoudjo et al., 2014;Arouna et al., 2016;Toko Imorou et al., 2019). The number of rows in the matrix indicates the number of land use units at time t 0 (1987), the number Y of columns in the matrix is the number of converted units at time t 1 (2020), and the diagonal contains the areas of the units that remained unchanged (Oloukoi et al., 2006;Avakoudjo et al., 2014;Ayena et al., 2017;Samba et al., 2021). The transformations are done from rows to columns (Toko Imorou et al., 2019;Samba et al., 2021). The cells of the matrix contain the value of a variable that moved from an initial class i to a final class j during the period from t 0 to t 1 . Column and row values represent the proportions of the areas occupied by each land use class at the corresponding time (Bamba et al., 2008;Ayena et al., 2017).
The annual deforestation rate was calculated for the period under consideration.

Confusion matrices and Kappa indices
The confusion matrix for the classification of the 1987 Landsat image reveals that 99.24% of the pixels in the class "Crop Lands, Pastures and Fallows" were well classified (Table 2). For the classes "Bare lands and buildings" and the class "Forests", all pixels were accurately classified. Only 0.51% of pixels in the class "Crop Lands, Pastures and Fallows" were confused with those in the class "Bare lands and Buildings". Overall, the results of the confusion matrix are acceptable.  In the class "Crop lands, Pastures and Fallows" ,99.46% of the pixels were well classified. For the classes "Bare lands and Buildings" and the class "Forests", all pixels were accurately classified. The percentage of pixels in the class "Crop lands, Pastures and Fallows" that were confused with those in the class "Bare lands and buildings" is 0.54%. Overall, the results of the confusion matrix are satisfactory.
The overall accuracy of the classification of the 2020 Sentinel-2 image is 99.61% while the Kappa index obtained is 0.99. The confusion matrix of the classification of this Sentinel-2 image is summarised in Table 4. According to the confusion matrix (Table 4), the classification was successful because the confusions between the different land use classes are low, with the exception of 0.81% of the pixels of the class "forests" which were confused with the pixels of the class "Crop lands, pastures, and fallows". In general, these results are acceptable because the South Talihya watershed is situated in an anthropized area where the clear distinction of land use classes is not very evident.

Pixel omission and commission errors
The percentages of pixel omission (Om) and commission (Com) errors in the classification of the 1987 Landsat image are reported in Table 5.  Table 5 (year 1987), the percentages of omission errors in the land use classes are low. The omission errors for the 'Fields, fallows and pastures' class represent only 0.76% while for the "forests" class, they represent 0.33%. For the class "Crop lands, fallows and pastures", the errors of the commission represent 3.22%, which is also low and not likely to taint the results of the classification. The results in Table 5 (year 2001), the percentages of omission errors in the land use classes are low. The errors of omission for the class "Crop lands, fallows and pastures" represent only 0.54%. For the class "Bare lands and buildings", the errors of the commission represent 6.67%, which is also low and not such as to taint the results of the classification. The results reported in Table 5 (year 2020) show that 0.81% of the pixels in the "forests" class were omitted while the commission rate is 1.11% for pixels in the "Crop lands, fallows, and pastures" class.

Dynamics of land cover between 1987 and 2020
The class "cropland, fallows, and pastures" was always the most dominant but decreased slightly between 1987 and 2001 before rising up in 2020. On the contrary, the "Bare land and buildings" class increased between 1987 and 2001 before dropping again in 2020. The class "forest" on its behalf is consistently decreasing from 1987 to 2020 (Figure 4). The values in the transition matrices below are presented as percentages of area. For ease of interpretation, the total area of the South Talihya watershed is estimated to be 560.5 km 2 , so 1% is approximately 5.605 km 2 .
The land use transition matrix in the South Talihya Watershed between 1987 and 2001) shows that bare lands and buildings increased from 10.48% in 1987 to 20.45% in 2001 (Table 6)  The annual deforestation rate calculated from the forest areas at date t0 (1987) and t1 (2020) is 1.60% for the South Talihya watershed.

Evaluation of the quality of classifications
The Maximum Likelihood classification algorithm gave good results. The preference of this classification method is due to its qualities related to its robustness in identifying classes that are spectrally close enough (Oszwald et al., 2010;Soro et al., 2013), which offers a good generalization capacity (Tente et al., 2019). It has given good results during the work of several authors (Soro et al., 2013) by the fact that it assigns each pixel to the class to which it has the highest probability of belonging (Katembera Ciza et al., 2015;Agbanou et al., 2018;Orekan et al., 2019;Manguiengha et al., 2019;Samba et al., 2021) by highlighting the standard margin of error between pixel values and those of different training sites (ROI) (Orekan et al., 2019).
The implementation of this supervised image classification with three classes for South Talihya gave very good Kappa index values (higher than 75%) based on the rating scale proposed by Landisand Koch (1977). According to Landis & Koch (1977) and Kundell & Polansky (2003), this index is "Excellent" when it is greater than or equal to 81%; "Good" when it is between 80% -61%; "Moderate" when it is between 60% -21%; "Poor" when it is between 20%-0%, and "Very Poor" when it is less than 0%. Pontius (2000); Oloukoi et al. (2006);Cabala Kaleba et al. (2017) demonstrated that a land-use study can be validated if the Kappa index is between 50% and 75%. Thus, the values of the Kappa index obtained in this study are in the range of 0.75 to 1 which is estimated as satisfactory in the framework of a maximum likelihood assisted classification for the tropical region (Sangne et al., 2015;Agbanou, 2018). Given the complexity of the landscape in these two watersheds, such a result can be explained by the quality of the images and the choice of thematic classes and training areas guided by the ground truth. In addition, the use of the 2020 Sentinel-2 image also contributed to the improvement of the classification quality.
According to Adjonou et al. (2019), the Sentinel-2 image shows better performance in the discrimination of land cover types. Although Cohen's Kappa index is widely used in works on land cover dynamics for classification validation (Barima et al., 2009;Kyale Koy et al., 2019), it is subject to many criticisms related to its misuse (Agbanou, 2018). This index is normally used in the case of systematic sampling, carried out randomly (Golden, 2006). However, despite the opinions tending to question the quality of the precision provided by Cohen's Kappa index, no index has yet been able to validly replace it in terms of evaluation and validation of the results obtained following the supervised classification of the different land use units (Agbanou, 2018).

Change In Land Cover and Deforestation Rates In The Watershed Between 1987 and 2020
The annual deforestation rate is 1.60% for the South Talihya watershed and is lower than the rate reported by Ndavaro et al. (2021) who found an annual deforestation rate of 2.2% in the Cool Highlands of Lubero. A slightly higher annual deforestation rate of 2.7% was obtained by Bweyaet al. These contrasting deforestation rates reveal that the global average hides strong disparities between particular regions (Debroux et al., 2007). But also, the divergence of data sources and processing methods, added to the difficulties of collection and harmonization, often lead to notable differences in the different evaluations of deforested areas (Herve et al., 2015). In addition, Yangambi Biosphere Reserve is a conservation area where human activities are limited. This may suggest that the annual deforestation rate is low compared to the South Talihya watershed where population density is high and land use activities are mainly agricultural and pastoral.
In addition, the analysis of the dynamics of change in land use classes in the South Talihya watershed has highlighted the various processes of evolution that occurred in the landscape during the period 1987-2020. These changes are mainly related to the regression of natural formations (deforestation process) such as dense forest formations in favor of Crop-fallows mosaics and bare lands and buildings areas that have also increased in the area. The factors most commonly cited as having strongly contributed to the reduction of forest and savanna areas include inappropriate agricultural practices, logging and charcoal production, and vegetation fires (Vyakuno, 2006;Adjonou et al., 2013). In this watershed, the manifestations of these factors are increasingly felt with the ever-increasing population growth in the area, thus increasing tenfold the need for croplands and increasing the pressure on natural resources (Vyakuno, 2006).

CONCLUSION
The objective of this paper was to study the dynamics of land use in the watershed of the South Talihya River. The results reveal that this dynamic of land use has led to the decrease of forest areas in favor of agricultural lands and fallows. Therefore, the rational management of natural resources in this agropastoral area and the restoration of forest ecosystems in this watershed with a high population density is necessary. This rational management will require a holistic approach to integrate a wide range of parameters and will guarantee sustainable development, a guarantee of a fulfilling life where the risk approach is integrated into the spatial planning process whose unit is the watershed.