Satellite images were pre-processed prior to thematic classification. This included color composition, image enhancement, and extraction of the study area. The colour composition was done to obtain a single multispectral image since Landsat images are designed as individual bands (Djagnikpo et al., 2016). Finally, the extraction of the study area was performed from the obtained multispectral images, following the boundaries of the Kimemi watershed.

To perform image classification, the supervised classification approach was chosen (Ali, 2016; Kabanyegeye et al., 2020; Pham & He, 2012; Salomon et al., 2020). In fact, this approach makes it possible to discriminate between different classes of land use and land cover (Ali, 2016). The Regions of Interest (ROI) were determined on the basis of geographical coordinates collected in the field and a visual interpretation of Google Earth images guided by knowledge of the study area. Three relevant classes were selected for the classification namely 1) the bare land and building (houses, roads and bare land), 2) woodland (tree plantations) and 3) grassland (lawns, cropland, etc.). The classification was based on the Maximum Likelihood algorithm. This algorithm is based on the statistics of the training areas, calculating the probability of a pixel belonging to a given class rather than to another. Pixels are assigned to the class for which the probability is highest (Djagnikpo et al., 2016; Girard & Girard, 2010).

The classification accuracy was assessed through the confusion matrix with the calculation of validation indices such as the overall accuracy and the Kappa index. This confusion matrix makes it possible to compare the predicted land use/land cover, i.e., that resulting from the classification, with the land use observed in the field (Girard & Girard, 2010; Justine, 2012). Kappa values below 50% indicate poor classification, while those between 50% and 75% and those above 75% indicate acceptable and excellent classification, respectively (Bogaert, Vranken, & Andre, 2014; Landis & Koch, 1977).

The spatiotemporal dynamic of the different land use/land cover classes was evaluated through the construction of the transition matrix, the calculation of the overall change rate, and the calculation of the spatial structure indices of the landscape. The transition matrix is a square matrix that allows to highlight in a condensed way the change of state undergone by the elements of a landscape during a given period (Bogaert et al., 2008; Djagnikpo et al., 2016). Three matrix (1987-2011, 2011-2015, 2015-2021) were produced by crossing the maps two by two for the periods considered. In addition, the overall change rate was also calculated to detect changes for each land use class. The spatiotemporal evolution of these classes was evaluated through the relationship between the same class at two different dates. This relationship made it possible to extract the stable, regressive and progressive areas of the focused class. The overall change rate (Tg) were calculated using Equation (1) proposed by the FAO in 1996 (Djagnikpo et al., 2016):

Tg = S 2 − S 1 S 1 × 100 (1)

With:

S₁: area of a class at date t1, in km²;

S₂: area of the same class at date t2, in km²;

Tg: overall change rate.

A value of Tg close to zero indicates the relative stability of the class, while positive values represent an increase in the area of the class during the period analysed. Negative values indicate the loss or regression of the area occupied by a class between the two dates.

The next step in the analysis was the calculation of spatial pattern indices. These indices are indicators of human impact on landscape configuration (Bogaert et al., 2008). In this study only considered the number of patches (n), the total area of patches (at), and the dominance index (D) computed after Bamba et al. (2008) as follows:

a t j = ∑ i = 1 n j a i j (2)

D J ( a ) = a max j a t j × 100 (3)

(2) Total patches area and (3) dominance index

With:

n_j: Total number of patches;

a_ij: area of patch (i) of class (j), in ha;

a_tj: total area of the class (j), in ha;

a_maxj: area of the largest patch of the class (j), in ha;

D_j(a): dominance index in %.

The spatial transformation processes during the study period were identified using the decision tree proposed by Bogaert, Ceulemans and Eysenrode (2004). This tree is based on the evolution of the number of patches, the total area and the total perimeter of patches of the land use class. The choice between fragmentation and dissection were based on a predefined reference value t = 0.5.

Table 2 presents the different values of the precision index found in this study. The analysis of the accuracy of the supervised classification of the images reveals overall accuracy values of 99.2%, 99.9%, 100%, 98.5% and Kappa index of 0.9871, 0.9981, 1, 0.9761, respectively, for the images of 1987, 2011, 2015, 2021.

These values indicate a statistically reliable discrimination of the different land cover classes.

The evolution of land use between 1987 and 2021 is globally characterised by an increase in the area occupied by the bare land and building class on the expense of the grassland class. The woodland class, on the other hand, has evolved with less pronounced variations (Figure 2).

Between 1987 and 2011, the class of bare land and building (Tg = 61.33%, from 18.5 km² to 29.8 km²) and the woodland class (Tg = 34.2%, from 9.3 km² to 12.5 km²) increased, while the grassland class decreased (Tg = −39.5%, 36.7 km² to 22.2 km²). The increase in the class of bare land and building is due to the conversion of 19.2% of grassland and 3.2% of woodland. The woodland class increased due to the conversion of 11.4% of grassland and 1.9% of bare soil (Table 3).

From 2011 to 2015, the class of bare land and building class increased at a rate 7.5% per year (29.8 km² to 38.7 km²), while that of grassland and woodland decreased from 22.2 km² to 13.9 km² (Tg = −37.3%) and from 12.5 km² to 11.9 km² (Tg = −4.9%), respectively. During this period, the increase observed for the bare land and building class is justified by the conversion of 9.4% of grassland and 5.7% of woodland (Table 3).

The dynamics observed during the period spanning from 2015 to 2021 is rather marked by small changes between classes. The overall change rates corresponding to each of the classes reveal a slight increase in the class of bare land and building (Tg = 2.1%), as well as in that of grassland (Tg = 1.9%). On the other hand, the woodland class decreased (−8.9%). In this period, 2.1% of the woodland class was converted to bare land and building and 4.3% to grassland. A portion (3.6%) changed from grassland to woodland, while 6.7% changed to bare land and building. Furthermore, 2.1% of the woodland area was converted to bare soil and built-up area (Table 3). Table 3 presents the transition matrix, and Figure 2 and Figure 3 illustrate the different changes in land use and land cover during the study period.

From 1987 to 2011, the most dominant process in the class of bare land and building is creation, explained by a simultaneous increase in the number of

patches (Figure 4(a)) and their total area (Figure 4(b)). On the other hand, for the grassland class, the spatial transformation process was dissection following an increase in the number of patches followed by a decrease in their surface but with t value (0.61) higher than 0.5. On the other hand, for the woodland class, the creation of new patches was also observed. From 2011 to 2015, the dominant process in the class of bare land and building remains the creation resulting from the simultaneous increase in the number of patches and their surface during this period. Similarly, for the grassland class, the dominant process remains dissection. However, the woodland class underwent the attrition process (Table 4) following the simultaneous decrease in the number of patches (Figure 4(a)) and their area (Figure 4(b)). Finally, between 2015 and 2021 patches of the bare land and building class continued to be created while the grassland class underwent dissection (t = 0.99) and the woodland class continued to undergo attrition (Table 4).

In general, the dominance values in the bare land and building class are high compared to other classes. The dominance increased from 75.89% in 1987 to 90.32% in 2011 reflecting the increase of the largest patch of the class in the landscape. This value stabilised until 2021 at around 91.33%. For the woodland class, the value of dominance also increased during the study period with values of 11.43%; 19.03%; 23.18%; 29.30% in 1987, 2011, 2015 and 2021 respectively. However, the dominance values of grassland class decreased drastically from 93.21% in 1987 to 10.55% in 2021.