Detection of scattered settlement in and around Babile Elephant Sanctuary and threats it pose using robust remote sensing method, Eastern Ethiopia.

doi:10.21203/rs.3.rs-3315400/v1

Download PDF

Research Article

Detection of scattered settlement in and around Babile Elephant Sanctuary and threats it pose using robust remote sensing method, Eastern Ethiopia.

https://doi.org/10.21203/rs.3.rs-3315400/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Illegal human activities occurring within protected areas (PA) are threats to the ecosystem and an annoyance for PA managers in many parts of the world including Ethiopia. This research aims to map scattered settlements in and surrounding Babile Elephant Sanctuary (BES) to analyse its impact on the endangered PA. Sentinel-2 L2 spectral data from dry and wet seasons were used to map settlements. A simple method of cloud masking based on multi-temporal thresholding that referenced a cloud-free dry season image was proposed to produce a cloud-free wet season image from several images. Then 19 features were derived from the seasonal images to be used as input for the random forest classifier. An integrated dataset was produced from the World Settlement Footprint and the new settlement map for analysis of settlement distribution. To see the variation in settlement density, cluster analysis was performed using Ordering Points to Identify the Clustering Structure (OPTICS) algorism. Our result was very closer to the World Settlement FootprintDataset The result showed that an area on the west of BES that covers only 18% of its boundary was settlement-free. If BES is to survive, relevant authorities should take immediate measures that curtain the movement of people into the sanctuary.

Cloud masking

Cluster analysis

Random Forest

Sanctuary

Scattered settlement

Sentinel

Illegal human activities occurring within protected areas(PA) are threats to the ecosystem and its management in many parts of the world. Fan et al. ( 2019) analyzed human development surrounding different types of terrestrial PAs, growth rate, and trend characteristics measured using remote sensing data. Based on the analysis of night time light from 1992 to 2013, they revealed human development inside and within 5 km buffer zones increased by two folds in type III PAs (natural monuments). The increase in the other types of PAs is even higher in the range of three to four folds. Using decadal spatial population data for Africa and Latin America from 1960 to 2000, Wittemyer et al. (2008) observed that population growth rates within 10 km of rural PAs were significantly higher than national rural averages and pose a threat by causing habitat disturbance.

Human settlement, agricultural encroachment, and overgrazing by livestock are considered the main threats to PAs in Ethiopia by several studies. Kumssa and Bekele (2014) describe the aggravated situation in Abjata-Shalla Lakes National Park as “the Park is congested with the homestead and currently, many new huts are being constructed”. A similar situation was observed in Senkele Swayne’s Hartebeest Sanctuary (Sultan et al., 2017). The problem of settlement intrusion is so severe that it became a political issue in protected areas (Boillat et al., 2018; Kelboro, 2012; Teressa, 2017; Tsegaye et al., 2017). The effect of these challenges can be a decline of endemic wildlife(Stephens et al., 2001), deterioration of habitat (Girma et al., 2018; Kiros & Bekele, 2021), or change of ecosystem by spreading invasive species (Seid & Kakiso, 2019). Lack of timely and accurate information is considered one of the causes intensifying such problems. Dudley et al. (2018) mention vulnerability assessments to be undertaken regarding protected areas, particularly for connectivity and corridors. This is especially true with rare and endangered species in developing countries, where wildlife managers have limited resources and information to formulate effective conservation strategies (Tefera, 2011).

An increasing number of studies have been carried out to extract human settlements, including urban and rural settlements using remote sensing data. However, the heterogeneity of rural landscapes with small and scattered houses intermingled with farms and trees poses a challenge and needs an innovative approach to extract data from remote sensing images. This contributes to the low accuracy of rural area mapping (Leyk & Uhl, 2018). Previous works trying to map scattered rural settlements are rare. They utilized either high-resolution images (Yang & Luo, 2012) or multi-source features (Hoffman-Hall et al., 2019; Marconcini et al., 2020; Tatem et al., 2004; Xu, 2021).

The heterogeneity of rural landscapes with small and scattered houses intermingled with farms and trees poses a challenge and needs an innovative approach to extract data from remote sensing images. Moderate spatial resolution remote sensing data offers a compromise between the fine spatial resolution of VHR data and the frequent temporal resolution of low spatial resolution data (Hoffman-Hall et al., 2019). The availability of preprocessed Level 2 Landsat and Sentinel-2 multispectral data is an advantage in this regard. In this study, scattered settlements are identified using multi-temporal sentinel-2 images. To make the wet season images appropriate for settlement identification clouds should be removed. The study applied a new method of cloud masking which is better than the scene classification mask of sentinel-2 L2 data. Several features derived from sentinel-2 imagery are compared and valuable spectral indexes and texture features are identified using existing methods.

This study attempts to identify single houses using multi-temporal optical remote sensing data. Specifically, it aims to: 1) Select a robust remote sensing method that can identify scattered settlements in a rural environment, and 2) Map current human settlement in and around Babile Elephant Sanctuary (BES) and analyze the threats it poses on the sanctuary. After identifying settlement pixels, it is important to explore the distribution and pattern of settlements in the area. This helps compare the intensity of settlement in different zones of the PA and identify the most affected areas and zones which are not settled yet. For this we applied spatial analysis including buffer analysis, spatial clustering and symmetric difference.

2.1. Study Area

The study area is a protected area in Ethiopia, designated as Babile Elephant Sanctuary (BES). BES is situated in Eastern Ethiopia between Oromia and Somali National Regional States. It is located between the latitudes of 08°02'29" N and 09°23'19" N and longitudes of 41°45'29" E and 43°22'21" E (Fig. 1). BES is the largest protected area in the country, covering an area of 8,388 km² (EWCA & WSD, 2010). The study area includes a 20 km buffer zone to compare the extent of settlement distribution in and around BES.

The sanctuary is characterized by arid and semi-arid climatic conditions with a bimodal rainfall pattern. The main land cover of the area is natural vegetation including open shrubland, dense shrubland, and woodlands. The vegetation of BES can be categorized into three major types: riverine forest, Acacia-Commiphora Woodland in lowlands and mid-highlands, and Bushland/scrubland (Zelalem Wodu, 2007). Including the African Elephants and Black Lions, the sanctuary harbours 30 species of mammals belonging to seven orders and 15 families. The sanctuary is one of the areas of unquestionable national significance to be included in the Ethiopian Protected Area System (Vreugdenhil et al., 2012). It is in category II of the International Union for Conservation of Nature (IUCN) Management category and is administered by Ethiopia Wildlife Conservation Authority (EWCA) (Gebremeskel Gizaw, 2021). It is increasingly suffering from the settlement, intensive agriculture, and livestock grazing. The area of BES is shared by 15 woredas of two regions. The population density in these woredas ranges from 21 in Degahamido of Somali Region to 711 in Haromaya of Oromia Region based on population predictions for 2017(CSA, 2013).

The study area is selected because it is one of the PAs in the country seriously affected by population pressure and settlement. Neil and Greengrass (2021) showed how uncontrolled immigration and settlement threatened BES with a four-fold increase in the number of houses from 2006 to 2017. Increasing demand for natural resources in the area like fodder, fuel wood, timber, water, and land for cultivation are reasons for human encroachment into the reserve area (Aschalew 2020; Reddy and Workneh 2014; Sintayehu & Kassaw, 2019). The size and quality of the sanctuary for the conservation of biodiversity have been reduced due to the huge immigration of farmers and their livestock who are permanently settled in and around the sanctuary (EWCA & WSD, 2010). Browsing animals, especially camels, are in direct competition for resources with the Elephants. BES has received the least attention for conservation since its establishment although it has been designated as the largest sanctuary for Elephants in the country (Belayneh et al., 2013).

2.2. Data source

Sentinel-2 level 2A Image was the main source of data for the settlement mapping. It is freely available at https://scihub.copernicus.eu/dhus/#/home. The Level-2A processing includes a scene classification and an atmospheric correction applied to Top-Of-Atmosphere (TOA) resulting in orthoimage, Bottom-Of-Atmosphere (BOA) corrected reflectance product with 10m, 20m, and 60m resolution (Sentinel Hub, 2022). Using multi-temporal data increases the accuracy of mapping scattered settlements (Hoffman-Hall et al., 2019; Ji et al., 2020; Marconcini et al., 2020; Muro et al., 2020; Xu, 2021). For this study, images from the two main seasons were used, i.e., wet season (August/September 2021) and dry season (March 2022). The total number of scenes used for this study was 14 (Table 1). The study area is covered by five tiles. The dry season images were used directly. But for the wet season images, cloud pixels were masked out from the main image (Table 1, column 3) and replaced by pixels from additional wet season images (Table 1, column 4).

Table 1

Images used for the study
Tile identifier	Dry season image date	Green season main image date	Green season additional image date
37PHK	06 Mar, 2022	28 Aug, 2021	08 Aug, 2021
38PKQ	06 Mar, 2022	28 Aug, 2021	12 Sep, 2021
37PHL	11 Mar, 2022	18 Aug, 2021	28 Aug, 2021
38PKR	08 Mar, 2022	28 Aug, 2021	08 Aug, 2021
38PLQ	16 Mar, 2022	30 Aug, 2021	-

2.3. Image Preprocessing and interpretation

Remotely sensed raw data generally contain flaws and defects. Removal of flaws through some methods is termed pre-processing methods and mainly includes geometric correction, radiometric correction, and atmospheric correction methods (Reddy, 2008). These corrections were not needed for Sentinel-2 level 2 data. Besides, Sentinel-2 L2A data are Ortho-rectified and UTM geocoded. Since sentinel-2 imagery comes with different spatial resolutions, resampling is needed. So, all images were resampled to 10 meters pixel size.

2.3.1. Cloud masking

Cloud detection is a preliminary step before processing optical remote sensing data (Hagolle et al., 2010). Clouds significantly alter the spectral signatures obtained from satellite data, which often leads to the false identification of land cover change (Tarrio et al., 2020). But, as the thermal band is still a key data source for cloud detection, cloud screening in Sentinel-2 is challenging (Tarrio et al., 2020). High omission of clouds and shadows was reported (Claverie et al., 2018; Tarrio et al., 2020). The data quality report (Enache, 2022) shows the overall accuracy of L2 scene classification varies for cloud versus clear pixels between 57.3 to 96.8%. The main problems were the over-detection of clouds over bright targets and the under-detection of semi-transparent clouds or cloud edges (Enache, 2022). This study applied a new method of cloud masking based on simple multi-temporal thresholding that referenced cloud-free dry season images. The advantage of using this method was described in the supplementary data. The threshold of 0.045 was found fairly good and applied for cloud masking in this study. The following equation is applied.

B1_w - B1_{d >} t_{−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−(1)}

Where B1_w and B1_d are reflectance values of the coastal aerosol band (B1) for the cloudy wet season image and the cloud-free dry season image respectively. t is a threshold that differentiates between cloudy and non-cloudy pixels.

2.3.2. Feature selection and training data preparation

For any modelling, the input data is the most important consideration. Remote sensing bands and their derivatives were used for feature classification. Textural indices derived from satellite imagery are useful to improve classification performance (Mellor et al., 2015). Co-occurrence measures were applied in ENVI software which was based on the work of Haralick, Dinstein, & Shanmugam (1973). All available texture measures were tried with the spectral indexes on a 3 by 3 window size. All features with promising results were selected. After recursive trials on the test area, 19 variables were selected to identify settlement structures (Table 2). These were based on 12 indexes, three texture metrics, and three-time variables. Feature importance was also analysed based on the random forest gain ratio. To prepare the training data, we used high-resolution remote sensing data to collect training samples. Training polygons were delineated for each class by overlaying Google earth image, true colour sentinel image and Normalized Difference Vegetation Index (NDVI) data. The samples were distributed all over the study area with more preference given to the overlapping area of tiles. This decreases workload by using the same training repeatedly. Areas with imagery dates before 2019 especially for settlement samples were avoided.

To avoid misallocation issues, samples were taken only from the middle of the class objects. However, this was not always possible as settlement areas were fragmented. The inclusion of the surrounding area outside the sanctuary was important because more compact settlements were found around this area. The training data quality was assessed by investigating the histogram (Lillesand et al., 2015).. When the histogram of a land cover class is uniform and unimodal for some feature, the training data can be considered of good quality. The training data collection produced 506 polygons containing 20,752 pixels for the training data, of which only 186 polygons or 1762 pixels (8.5%) were for the settlement class.

Table 2

Features used as inputs for classification
No	Index/variable	Equation	Reference	Seasonal variable	Texture metrics
1	Built-Up Areas Index (BAI)	(B2 – B8)/ (B2 + B8)	(Mhangara et al., 2011)	Dry and wet	-
2	Brightness Index- BI	[(B4² + B3²)/2]^0.5	(Escadafal, 1989)	dry	-
3	Bare soil index (BSI)	[(B11 + B4)- (B8 + B2)]/ [(B11 + B4)+ (B8 + B2)]	(Chen et al., 2004)	wet	contrast
4	Band Ratio for Built-up Area (BRBA)	B4 / B11	(Waqar et al., 2012)	dry	variance
5	Brightness index (BI3)	(B4 + B3 + B2)/3	-	Seasonal difference	Mean of dry season index
6	Brightness index (BI4)	(B2 + B3 + B4 + B8)/4	(Zhou et al., 2009)	dry	-
7	Modified Bare Soil Index (MBSI)	2(B4-B3)/ (B4 + B3-2)	(Zhang et al., 2018)	wet	-
8	Normalized Difference Green Blue (NDGB)	(B3 – B2)/ (B3 + B2)	(Marconcini et al., 2020)	Dry and wet	-
9	Normalized Difference Red Blue (NDRB	(B4 – B2)/ (B4 + B2)	(Marconcini et al., 2020)	dry	-
10	B8-B1 index	(B8– B1)/ (B1 + B8)	-	dry	-
11	Tasseled Cap Wetness (TCW)	(B2∗0.2578)+(B3∗0.2305)+ (B4∗0.0883)+(B8∗0.1071)- (B11∗0.7611)-(B12∗ 0.5308)	(Shi & Xu, 2019)	Wet and seasonal difference	-
12	B11	Raw sentinel SWIR band		Dry season and seasonal difference	-

2.3.3. Classification method and accuracy assessment

Although the target of this study was to identify settlement areas, multiple classes were used based on a pre-experiment test. More classes are necessary to improve classification accuracy by encompassing the spectral variability of a landscape (Adams & Gillespie, 2006). Six-class were identified (two classes of settlement and four classes of non-settlement). The settlement classes are bright roofs and dark roofs. Non-settlement classes comprise roads, bare land, seasonal vegetation and permanent vegetation. Because it is successful in several studies, the random forest algorithm (RF) was applied. Two parameters are important in the RF algorithm. The number of trees is usually set between 100 and 1000 (Rodriguez-galiano et al., 2012). Tree depth (the maximum number of splits in a tree) is usually set to the square root of the total number of input variables for classification (Belgiu & Drăgu, 2016). Hoffman-Hall et al.(2019) applied 500 trees and 9 splits in a similar study. In this study, the number of trees was set to 120, and the tree was based on a pre-test result. The default tree depth of 30 was maintained in ArcGIS. The classification was done for each tile separately mainly because mosaicking 5 tiles of sentinel data for 13 bands and 10m resolution is computationally intensive. After classification, the results of all tiles were merged and reclassified into binary classes.

Regarding the accuracy assessment, the result was compared with World Settlement Footprint (WSF) 2019, which is a 10m resolution global dataset (Marconcini et al., 2020) and available at https://download.geoservice.dlr.de/WSF2019/. WSF uses multi-temporal Sentinel-1 (S1) radar and Landsat-8 optical imagery as an input and Support Vector Machines (SVM) as a classifier. It also incorporates existing settlement datasets in the -post-classification steps. It was selected for its high accuracy. The agreement and disagreement areas were identified and compared.

Table 3

Number of samples for accuracy assessment
Density Class	Settlement density(pixels/km²)	Area in class- km² (%)	Number of samples for accuracy
1	0	131559 (69)	114
2	1–50	3964 (21)	182
3	51–500	1718 (9)	164
4	> 500	351(2)	87
	Total		545

A formal accuracy assessment was also done. Sampling design, response design, and error analysis are important considerations in accuracy assessment (Congalton & Green, 2019; Jensen, 2015). The sampling design is stratified random sampling. Stratification was based on the settlement density of the resulting map. Settlement pixels were counted in 1km square grids and the study area was divided into 4 classes based on settlement density. This served as strata for sampling. The total number of samples was 545. The number of samples for each stratum was allocated as mentioned in Table 3. More samples were allocated to the middle classes (2nd and 3rd ) to emphasize the scattered settlements as it was the focus of this study. The sampling unit was 3 by 3 pixels of the resulting map to avoid the transfer of misallocation error. when there was any settlement structure in a 30 by 30m grid, it was considered as reference settlement data. Otherwise, it was recorded as non-settlement. Reference samples were collected from Google earth and using GPS. Overall accuracy, F-score, omission error, and commission error were used as accuracy measures. The F-score (Pérez-Hoyos; A. Udías & Rembold, 2020) is more meaningful for imbalanced classification. It is the harmonic mean of the user’s and producer’s accuracy.

F-score= (2*UA*PA)/ (UA + PA) --------------(2)

Where UA and PA are user’s accuracy and producer’s accuracy respectively. The F-score ranges between 0 and 1. The higher the value of the F-score the better the accuracy.

2.3.4. Settlement map integration method

The WSF map and the classified map of this study were integrated and used for subsequent analysis. First, the settlement pixels of WSF were masked. When the non-settlement area in WSF was mapped as a settlement in the study, the following threshold expressions were applied to remove the commission error from the classified map: Reflectance of B6 in the wet season > 0.22, BAI for wet season > -0.6 and NDGB < 0.9. These threshold values were joined by the Boolean ‘AND’ to remove less bright objects as well as contrasts of vegetation and bare land which were confused with the settlement.

2.4. Spatial analysis of settlement distribution

To see the variation in settlement density, cluster analysis was performed. First, the raster data was changed to point. Then the points were changed to clusters with the Ordering Points To Identify the Clustering Structure (OPTICS) algorithm using the Density-Based Spatial Clustering with the Application of Noise (DBSCAN) package (Hahsler et al., 2019). Density-based clustering works by grouping regions of high density and separating them from regions of low density. This helps identify coherent, homogeneous regions in the data set (Pattnaik, 2020). The minimum number of points (minPts) was set to 3 and the neighbourhood distance (eps_cl) to 500 meters. The polygon enclosing each cluster was drawn using the concave hull algorithm (Gombin et al., 2020) in R packages. The second iteration of DBSCAN was done on polygons with an area of less than 100 m². Lastly, the concave hull algorithm was performed on all clusters and the geometry was cleaned by dissolving overlapping boundaries. Settlement density was calculated by dividing the number of settlement points contained in a polygon by its area.

Comparison of settlement by proximity was also conducted using buffer zones. This study includes an 11,579 km² buffer area on the outer boundary of BES for comparison of settlement inside and outside the PA. The BES inside the area was 8388 km². Concentric buffers were drawn at 10 km intervals inside and outside the BES for comparison. The number of settlement clusters and their density were compared in these proximity zones.

Furthermore, it will be helpful for future decision-making if the location of extended settlement-free areas in BES is mapped. To map non-settlement area, first, the settlement polygons inside BES were buffered by 1 km. Then, the non-settlement area was identified by calculating the symmetric difference from the BES boundary. The resulting polygons were ranked by their area to identify where extended settlement-free areas were located.

3.1. Settlement Mapping

The image was classified into binary (settlement and non-settlement) classes following the procedure as described in the methods section. The classification was performed using the RF algorithm for the five tiles independently because of resource limitations to processing such large data. Feature importance was analyzed based on the random forest gain ratio and found that all features have significant importance.

As can be seen in Fig. 2, the contribution of the last-ranked features was about a third of that of the feature ranked first. But the contribution of features extracted from seasonal differences was not as expected. All seasonal difference features were ranked in the bottom four. On the other hand, wet season features and texture measures were on the highest rank. Another important observation from Fig. 2 was that features derived from the visible spectra, especially B2 (blue-reflectance), have very high importance in identifying anthropogenic features. After classification, the validity of the produced map was examined. It was compared with existing data followed by a formal accuracy assessment.

3.2. Comparison of the produced map with an existing dataset

The resulting settlement map was compared with the WSF (Marconcini et al., 2020) dataset. As can be seen from Table 4, the general appearance of settlement in this study was similar to WSF. About 740, 000 pixels were settlement in WSF, while the respective settlement pixels in this study were about 780, 000, indicating only a 5% difference between the two datasets. Therefore, the total number of settlement pixels in the study area was comparable. Figure 3 shows the agreement map.

Upon closer observation, however, there was more variance between the two datasets. When we see pixel by pixel agreement, there was more disagreement than agreement as shown in Table 4. The magnified view-A showed an agreement area. This was typically true for the center of towns and cities. View B showed an area that was mapped as a settlement in WSF but not in this study. This was mainly on the outskirts of towns. View C indicated an area mapped as a settlement in this study, contrary to WSF. This was a weakness of this work where the vibrant road surface was confused with the settlement.

Table 4

Agreement matrix of the two maps (WSF and this study)
Property	Number of pixels(x10⁶)	Percentage of the study area
Settlement in both	0.3016	0.157%
Settlement in WSF but not in this study	0.4399	0.230%
Settlement in this study but not in WSF	0.4832	0.252%
Non-Settlement in both	190.4341	99.361

3.3. Accuracy assessment

Accuracy assessment was applied using stratified random samples using google earth imagery and ground samples as reference. The number of samples was 545. Samples were evaluated in 30m-by-30m grids both in the map and reference google earth images. when there was any settlement structure in this grid, the sample was considered as a settlement, otherwise, it was recorded as non-settlement. The results of the accuracy assessment were presented in Table 5. The results showed that both datasets have a high overall accuracy (above 95%). But the overall accuracy didn’t tell much as the non-settlement area is very large and even high acuracy could be obtained by chance. The omission and commission errors were the actual measures and the F-score was a combination of the two measures. In this regard, WSF performs a little better.

Table 5

Accuracy measures.
Accuracy measure	This study	WSF
Overall accuracy	95.8%	97.2%
Omission error	14.3%	17.9%
Commission error	44.2%	30.3%
F-score	0.676	0.754

The map produced by this classification had a higher commission error (44.2%) than WSF (30.3%). This was in agreement with the discussion in the previous section. Some bright surfaces and main roads were confused with roofs. Concerning omission error, the dataset of this study performs better. It was observed that the missed locations were either isolated very small single houses or roofs made of very dim material very similar to the background soil. An integrated dataset was produced from the WSF dataset and the settlement map produced in this study (Fig. 4). For the non-settlement area in WSF, additional constraints were applied based on the threshold to remove the commission error from the classified map. The subsequent analysis was based on the integrated map.

3.4. Spatial Analysis of Settlement Distribution

The final settlement map with a spatial resolution of 10m is shown in Fig. 4. The number of settlement pixels in the study area was 750,755, of which 57,191 were in the boundary of BES. Although the dense settlements were surrounding the sanctuary, there were scattered settlements all over the study area including the center of the sanctuary. Neil & Greengrass (2021) reported 50,066 houses inside BES in 2017, which was based on a manual count of settlement structures from google earth image. This figure was comparable with our result, despite the difference in the method (visual count versus machine-based classification), unit of measurement (houses versus pixels), and reference year (2017 versus 2021). It should be noted that the number of pixels was not equivalent to the number of houses. The settlement pixel (100 m²) often contains more than one house (usually 2–3) especially when the density was higher as in the case of towns and villages.

3.4.1. Settlement clusters

The settlement pixels were converted to points and clustering was performed with the OPTICS algorithm by setting minimum points to 3 and neighbourhood distance to 500 meters. 750,755 settlement points were grouped in 8,235 clusters and 10,391 noise points (i.e., only 1.4% of the settlement points were not included in the clusters). The density of clusters was calculated by dividing the number of points in each polygon by its area. 6041(73%) clusters were less than 2.5 ha, indicating that most of the clusters were smaller in area. Besides, most of the clusters were also in the sparse class (Fig. 5).

3.4.2. Settlement distribution by proximity

Figure 6 shows the distribution of settlement clusters and their density. Most of the denser clusters were found in the northern part of the study area outside the boundary of BES. But a significant portion of the larger clusters was in the middle of BES, especially along the Babile-Fik Road. The study area was divided into four concentric zones using a 10 km buffer distance (two inside and another two outside the boundary). Then, the number of settlement clusters for the three zones, i.e., zones 1,2,3(Fig. 6) was compared by area and density.

The settlement density and size are summarized in Table 6. Except for clusters that were both dense and large, the number of clusters inside BES (zone 1 and 2) was close to that of outside the BES (zone 3). In some cases, for example, for the sparse medium area class, the number of clusters inside (88) was higher than that of outside (50). This showed that there were not only scattered settlements but also villages and towns inside the sanctuary. This can be considered a human intrusion and a serious threat to the sanctuary. Previous studies also showed that encroachment of human settlement in and around PAs was one of the most important threats for protected areas in eastern Ethiopia and BES was the most threatened one (Wale et al., 2017).

Figure 7 is a summary of Table 6, showing the trend of settlement distribution from the center to the outside of the BES. The slope of the trend lines from zone 1 to zone 2 was negative for the large and medium-sized clusters (> 2.5 ha) and it was positive from zone 2 to zone 3(i.e., from the periphery to the outer area). This implies that settlers were not afraid of the laws prohibiting them from entering the BES, rather they went directly to the place they preferred for settlement. Similar tendencies were seen in Nech Sar National Park(Wana et al., 2021).

Table 6

Number of settlement points by density and size (area) in the 3 proximity zones
Cluster area (ha)	Cluster Density (set. Points/ha)	Zone 1	Zone 2	Zone 3
Number of clusters by area and density
< 2.5	< 0.5	1	0	1
< 2.5	0.5-10.47	428	714	1481
< 2.5	> 10.47	244	440	990
Total small area (< 2.5 ha) clusters		673	1154	2472
2.5–500	< 0.5	88	42	50
2.5–500	0.5-10.47	341	341	741
2.5–500	> 10.47	2	2	31
Total medium area (2.5–500 ha) clusters		431	385	822
> 500	< 0.5	4	3	0
> 500	0.5-10.47	0	0	12
> 500	> 10.47	0	0	1
Total large area (> 500 ha) clusters		4	3	13
Total number of clusters		1108	1542	3307
Total Number of clustered points		31,326	22,032	277,917
Number of Noise points out of cluster		1699	2134	3677
Total number of settlement points		33,025	24,166	281,594

Zones 1, 2, and 3 are 10km concentric buffer areas from the center to the outside respectively. Their area is 4615, 4170 and 5393 square km respectively.

3.4.3. Non-settlement area

By using only a 1 km buffer around the settlement clusters, we tried to identify the non-settlement areas. As can be seen from Fig. 8, the largest extended settlement-free area was found in the west of BES. This relatively extended non-settlement area covered 1,591 km², which was only 18% of the total area of BES. When we consider the annual home range limit of the Elephants as the confluence of the Erer and Gobele rivers (Demeke et al., 2012), the effective area was reduced to 1262 Km².

Another major insight that can be inferred from Fig. 8 was that the main settlements were associated with roads and rivers. Settlements followed the roads indicating the high need of the society for development of infrastructure. Settlements were also associated with river routes. In a dry environment, water is an important resource for irrigation agriculture and domestic supply. The river lines may be the main source of Human-Wildlife conflict as the elephants should get water and feed from there (Aschalew, 2020; Neil & Greengrass, 2021; Reddy & Workneh, 2014).

Moreover, additional settlements were continuing to be established. From field observation (Fig. 9) and communication with the BES warden, the author noticed that no measures were being taken to avert the problem. Rangers feared moving to the interior of the protected area due to life threats from settlers. If the situation continues as it is, only a few years will remain to destroy BES.

A re-demarcation of BES was made recently by abandoning some areas in the north and compensating it in the south (EWCA & WSD, 2010). This was a reasonable decision given the intense population pressure from the North. But it cannot be reapplied in the future. Because it is not relevant to delineate an area outside of the natural range for the elephants, and more vigorous population pressure is coming from the south to north direction. So, the sanctuary manager should think of a different solution if BES is to sustain itself in the sea of challenges. This can be related to dealing with the surrounding community and local administrations that establish ownership positions.

Timely and accurate information is critically important in many areas for decision-making. Remote sensing is a less expensive and effective way of getting valuable and accurate information, especially in a large area. This study aimed to identify better methods of mapping scattered settlements and using the data in a case to analyze the problem of a protected area in Babile Elephant Sanctuary (BES), Eastern Ethiopia. We tried to map settlements using multi-temporal sentinel-2 L2 spectral data from two seasons. The dry season images were used directly. But for the wet season images, we applied a new method of cloud masking based on simple multi-temporal thresholding that referenced cloud-free dry season images. Nineteen variables were derived that comprised twelve spectral indexes, three texture metrics, and three-time variables from sentinel-2 images. The random forest classifier was used to classify the variables into six classes. Then, a binary (settlement and non-settlement) map was produced by reclassifying the six classes. The result was compared with World Settlement Footprint (WSF) 2019. Accuracy assessment was also done for both WSF and the new settlement map of the study area. A stratified random sampling with a sample size of 545 was performed and evaluated with F-score, omission error, and commission error. The map produced by this classification had a higher commission error (44.2%) than WSF (30.3%). But omission error of the new map (14.3%) was better than WSF (17.9%). An integrated map of the two datasets was used for settlement distribution analysis.

The number of settlement pixels in the study area was 750,755, of which 57,191 were in the boundary of BES. This was very close to a result based on a manual count of settlement structures from google earth images in a previous study. Cluster analysis was done using the OPTICS algorithm to map the settlement polygons by density. The polygons were used to compare the settlement intensity by proximity from inside to outside. The results showed that the number of large sizes (> 2.5 ha) settlements was higher in the centre of BES (435) than in the periphery (388). This implies the problem was not a border effect. By drawing a 1 km buffer around settlement polygons and calculating the symmetric difference with the boundary of BES, the non-settlement area was mapped. The result showed that the largest extended settlement-free area was found in the west of BES and covers only 18% of the total area of BES.

This study concluded that more than 80% of BES was already threatened by settlement. If the situation continues unabated, BES will disappear in a few years. EWCA (the authority in charge of administering the BES) should take immediate measures that limit the movement of people into the BES. The measures may include law enforcement, community dialogue, and utilization of resources for tourism and education. The insights gained from this study may also be used to develop a monitoring tool that helps to detect and report any LULC change in and around BES in near-real time before it becomes too late to take action. Although the result of this study has a higher commission error as compared to a very accurate WSF dataset, the study demonstrated that sentinel-2 images can be used to map scattered settlements with reasonable accuracy. This study identified important features for settlement mapping from a list of many spectral indexes and showed the importance of multi-temporal images. The study also presented a new cloud masking method based on multi-temporal thresholding which was accurate and much better than the L2 scene classification results.

It is hard to identify houses made of wood/grass, mud, and other natural materials with only optical imagery. Excluding such structures will make the mapping of settlement incomplete. A further study could try to improve the mapping of scattered settlements by using more advanced machine learning techniques like deep learning. Incorporating additional data other than spectral information like radar and auxiliary data may also improve accuracy. Another drawback of this study is that the attempt to identify the driving factors of settlement and its impact analysis does not show the full picture because of the scope of this study.

Adams, J. B., & Gillespie, A. R. (2006). Remote Sensing of Landscapes with Spectral Images. Cambridge: CAMBRIDGE UN I VERSITY PRESS.
Aschalew, A. (2020). Human-Elephant Conflicts in and Around Reserve Areas . A case of Babile Elephant Sanctuary of East. Solid State Technology, 63(6), 17866–17883.
Belayneh, A., Yohannes, T., & Worku, A. (2013). Recurrent and extensive forest fire incidence in the Bale Mountains National Park (BMNP), Ethiopia: Extent, Cause and Consequences. International Journal of Environmental Sciences, 2(1), 29–39.
Belgiu, M., & Drăgu, L. (2016). Random forest in remote sensing: A review of applications and future directions. ISPRS Journal of Photogrammetry and Remote Sensing, 114, 24–31. https://doi.org/10.1016/j.isprsjprs.2016.01.011
Boillat, S., Gerber, J. D., Oberlack, C., Zaehringer, J. G., Speranza, C. I., & Rist, S. (2018). Distant interactions, power, and environmental justice in protected area governance: A telecoupling perspective. Sustainability (Switzerland), 10(11). https://doi.org/10.3390/su10113954
Chen, W., Liu, L., Zhang, C., Wang, Jihua, Wang, Jindi, & Pan, Y. (2004). Monitoring the seasonal bare soil areas in Beijing using multi-temporal TM images. International Geoscience and Remote Sensing Symposium (IGARSS), 5(3), 3379–3382. https://doi.org/10.1109/igarss.2004.1370429
Claverie, M., Ju, J., Masek, J. G., Dungan, J. L., Vermote, E. F., Roger, J., … Justice, C. (2018). The Harmonized Landsat and Sentinel-2 surface reflectance data set. Remote Sensing of Environment, 219(October), 145–161. https://doi.org/10.1016/j.rse.2018.09.002
Congalton, R. G., & Green, K. (2019). Assessing the Accuracy of Remotely Sensed Data Principles and Practices (Third Edit). Taylor & Francis Group.
CSA. (2013). Federal Demographic Republic of Population Projection of Ethiopia from 2014 – 2017: Population Projection of Ethiopia for All Regions At Woreda Level from 2014-2017. 1–118.
Demeke, Y., Renfree, M. B., & Short, R. V. (2012). Historical range and movements of the Elephants in Babile Elephant Sanctuary , Ethiopia. African Journal of Ecology, (50), 439–445.
Dudley, N., Hockings, M., Stolton, S., Amend, T., Badola, R., Bianco, M., … Zhang, Y. (2018). Priorities for protected area research. Parks, 24(1), 35–50. https://doi.org/10.2305/IUCN.CH.2018.PARKS-24-1ND.en
Enache, S. (2022). Data Quality Report Sentinel-2 MSI L2A May 2022. Retrieved July 13, 2022, from Sentinel Online -Data Product Quality Reports website: https://sentinels.copernicus.eu/web/sentinel/data-product-quality-reports
Escadafal, R. (1989). Remote sensing of arid soil surface color with Landsat thematic mapper. Advances in Space Research, 9(1), 159–163. https://doi.org/10.1016/0273-1177(89)90481-X
EWCA, & WSD. (2010). Management Plan of Babile Elephant Sanctuary. Addis Ababa, Ethiopia. https://doi.org/10.20935/al788
Fan, L., Zhao, J., Wang, Y., Ren, Z., Zhang, H., & Guo, X. (2019). Assessment of night-time lighting for global terrestrial protected and wilderness areas. Remote Sensing, 11(22), 1–20. https://doi.org/10.3390/rs11222699
Gebremeskel Gizaw. (2021). Final data of ETHIOPIA PAs. Retrieved March 25, 2022, from Ethiopian Biodiversity Institute Geoportal website: http://196.188.40.146/documents/706
Girma, Z., Chuyong, G., & Mamo, Y. (2018). Impact of Livestock Encroachments and Tree Removal on Populations of Mountain Nyala and Menelik’s Bushbuck in Arsi Mountains National Park, Ethiopia. International Journal of Ecology, 2018. https://doi.org/10.1155/2018/5193460
Gombin, J., Vaidyanathan, R., & Agafonkin, V. (2020). concaveman: A Very Fast 2D Concave Hull Algorithm.
Hagolle, O., Huc, M., Pascual, D. V., & Dedieu, G. (2010). A multi-temporal method for cloud detection, applied to FORMOSAT-2, VENµS, LANDSAT and SENTINEL-2 images. Remote Sensing of Environment, 114(8), 1747–1755. https://doi.org/https://doi.org/10.1016/j.rse.2010.03.002
Hahsler, M., Piekenbrock, M., & Doran, D. (2019). dbscan: Fast density-based clustering with R. Journal of Statistical Software, 91, 1–30.
Haralick, R. M., Dinstein, I., & Shanmugam, K. (1973). Textural Features for Image Classification. IEEE Transactions on Systems, Man and Cybernetics, SMC-3(6), 610–621. https://doi.org/10.1109/TSMC.1973.4309314
Hoffman-Hall, A., Loboda, T. V., Hall, J. V., Carroll, M. L., & Chen, D. (2019). Mapping remote rural settlements at 30 m spatial resolution using geospatial data-fusion. Remote Sensing of Environment, 233(August), 111386. https://doi.org/10.1016/j.rse.2019.111386
Jensen, J. R. (2015). INTRODUCTORY DIGITAL IMAGE PROCESSING A Remote Sensing Perspective (4th Editio). Pearson Education, Inc.
Ji, H., Li, X., Wei, X., Liu, W., Zhang, L., & Wang, L. (2020). Mapping 10-m resolution rural settlements using multi-source remote sensing datasets with the google earth engine platform. Remote Sensing, 12(17), 1–23. https://doi.org/10.3390/rs12172832
Kelboro, G. (2012). Contesting the National Park theorem? Governance and land use in Nech Sar National Park, Ethiopia. Bonn.
Kiros, S., & Bekele, A. (2021). Assessment of conservation challenges in and around Gibe Sheleko National Park, southwestern Ethiopia. Global Ecology and Conservation, 32(July), e01912. https://doi.org/10.1016/j.gecco.2021.e01912
Kumssa, T., & Bekele, A. (2014). Attitude and perceptions of local residents toward the protected area of Abijata-Shalla Lakes National Park (ASLNP), Ethiopia. Journal of Ecosystem & Ecography, 4(1), 1.
Leyk, S., & Uhl, J. H. (2018). Data Descriptor : HISDAC-US , historical settlement data compilation for the conterminous United States over 200 years. SCIENTIFIC DATA, 5, 1–14. https://doi.org/10.1038/sdata.2018.175
Lillesand, T. M., Kiefer, R. W., & Chipman, J. W. (2015). REMOTE SENSING AND IMAGE INTERPRETATION (Seventh Ed). Hoboken: John Wiley & Sons, Inc.
Marconcini, M., Metz-marconcini, A., Üreyen, S., Palacios-lopez, D., Hanke, W., Bachofer, F., … Strano, E. (2020). Outlining where humans live , the World Settlement Footprint 2015. Scientific Data, (7), 242. https://doi.org/10.1038/s41597-020-00580-5
Mellor, A., Boukir, S., Haywood, A., & Jones, S. (2015). Exploring issues of training data imbalance and mislabelling on random forest performance for large area land cover classification using the ensemble margin. ISPRS Journal of Photogrammetry and Remote Sensing, 105, 155–168. https://doi.org/10.1016/j.isprsjprs.2015.03.014
Mhangara, P., Odindi, J., Kleyn, L., & Remas, H. (2011). Road extraction using object oriented classification. Visualisation Technical, 45–50.
Muro, J., Zurita-Arthos, L., Jara, J., Calderón, E., Resl, R., Rienow, A., & Graw, V. (2020). Earth observation for settlement mapping of Amazonian indigenous populations to support SDG7. Resources, 9(8). https://doi.org/10.3390/RESOURCES9080097
Neil, E., & Greengrass, E. (2021). Illegal settlement in the Babile Elephant Sanctuary is threatening the resident elephant population. Oryx, 1–8. https://doi.org/10.1017/S0030605320001088
Pattnaik, A. (2020). Geospatial Clustering: Kinds and Uses. Retrieved October 20, 2022, from Towards Data Science website: https://towardsdatascience.com/geospatial-clustering-kinds-and-uses-9aef7601f386
Pérez-Hoyos; A. Udías, A., & Rembold, F. (2020). Integrating multiple land cover maps through a multi-criteria analysis to improve agricultural monitoring in Africa. Int J Appl Earth Obs Geoinformation, 88(102064). https://doi.org/10.1016/j.jag.2020.102064
Reddy, M. A. (2008). Remote Sensing and Geographical Information Systems (Third). BS Publications.
Reddy, R. U., & Workneh, S. (2014). Conflicts between the conservation of Elephant and Human activities : In the case of Babile Elephant Sanctuary ( BES ), Ethiopia. Jouirnal of Geography and Regional Planning, 7(April), 25–29. https://doi.org/10.5897/JGRP12.036
Rodriguez-galiano, V. F., Ghimire, B., Rogan, J., Chica-olmo, M., & Rigol-sanchez, J. P. (2012). An assessment of the effectiveness of a random forest classifier for land-cover classification. ISPRS Journal of Photogrammetry and Remote Sensing, 67, 93–104. https://doi.org/10.1016/j.isprsjprs.2011.11.002
Seid, M., & Kakiso, T. (2019). Critical solutions for critical problems: Threats to Sustainable use and Management of Nech Sar National Park (NSNP) in Ethiopia. African Journal of Hospitality, Tourism and Leisure, 8(1), 1–20. https://doi.org/10.20546/ijcrar.2019.701.002
Sentinel Hub. (2022). Sentinel Hub. Retrieved April 9, 2022, from https://docs.sentinel-hub.com/api/latest/data/sentinel-2-l2a/
Shi, T., & Xu, H. (2019). Derivation of Tasseled Cap Transformation Coefficients for Sentinel-2 MSI At-Sensor Reflectance Data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 1–11. https://doi.org/10.1109/JSTARS.2019.2938388
Stephens, P. A., D’Sa, C. A., Sillero-Zubiri, C., & Leader-Williams, N. (2001). Impact of livestock and settlement on the large mammalian wildlife of Bale Mountains National Park, southern Ethiopia. Biological Conservation, 100(3), 307–322. https://doi.org/10.1016/S0006-3207(01)00035-0
Sultan, M., Amano, T., Mangesha, G., & Gure, A. (2017). Assessment of Attitude and Perception of Local Community toward Protected Area : The Case of Senkele Swayne ’ s Hartebeest. International Journal of African and Asian Studies, 31, 16–24.
Tarrio, K., Tang, X., Masek, J. G., Claverie, M., Ju, J., Qiu, S., … Woodcock, C. E. (2020). Comparison of cloud detection algorithms for Sentinel-2 imagery. Science of Remote Sensing, 2(September), 100010. https://doi.org/10.1016/j.srs.2020.100010
Tatem, A. J., Noor, A. M., & Hay, S. I. (2004). Defining approaches to settlement mapping for public health management in Kenya using medium spatial resolution satellite imagery. Remote Sensing of Environment, 93(1–2), 42–52. https://doi.org/10.1016/j.rse.2004.06.014
Tefera, M. (2011). Wildlife in Ethiopia: endemic large mammals. World Journal of Zoology, 6(2), 108–116.
Teressa, D. K. (2017). Ethiopia: Changes from “People out Approach” Protected Area Management to Participatory Protected Area Management? Insight from Ethiopian Protected Areas. IOSR Journal of Environmental Science, Toxicology and Food Technology, 11(2), 49–55. https://doi.org/10.9790/2402-1102014955
Tsegaye, G., Dondeyne, S., Lemenih, M., Marye, A., Nyssen, J., Deckers, J. A., & Maertens, M. (2017). ‘Facing conservation’ or ‘conservation with a human face’? People–park interactions in southern Ethiopia. Journal of Eastern African Studies, 11(2), 290–309. https://doi.org/10.1080/17531055.2017.1327167
Vreugdenhil, D., Vreugdenhil, A. D., Tilahun, T., Shimelis, A., & Tefera, Z. (2012). Gap Analysis of the Protected Areas System of Ethiopia.
Wale, M., Kassie, A., Mulualem, G., Tesfahunegny, W., & Assefa, A. (2017). Wildlife Threats and Their Relative Severity of Eastern Ethiopia Protected Areas. Ecology and Evolutionary Biology, 2(4), 59. https://doi.org/10.11648/j.eeb.20170204.12
Wana, D., Smith, S. W., Coleman, K. E. R., & Speed, J. D. M. (2021). Proximity to high densities of pastoral settlements reduces grassland regrowth in a protected tropical savanna. Biotropica, 53(5), 1316–1327. https://doi.org/10.1111/btp.12977
Waqar, M. M., Mirza, J. F., Mumtaz, R., & Hussain, E. (2012). Development of New Indices for Extraction of Built-Up Area & Bare Soil. Open Access Scientific Reports, 1(1), 1–4. https://doi.org/10.4172/scientificreports.13
Wittemyer, G., Elsen, P., Bean, W. T., Burton, A. C. O., & Brashares, J. S. (2008). Accelerated human population growth at protected area edges. Science, 321(5885), 123–126.
Xu, R. (2021). Mapping rural settlements from landsat and sentinel time series by integrating pixel-and object-based methods. Land, 10(3), 1–18. https://doi.org/10.3390/land10030244
Yang, C., & Luo, Z. (2012). Extracting rural settlement information from quickbird images. Key Engineering Materials, 500, 450–457. https://doi.org/10.4028/www.scientific.net/KEM.500.450
Zelalem Wodu. (2007). Elephant and Anthropogenic Impacts on Woody Plant Species in Babile Elephant Sanctuary, Eastern Ethiopia. ADDIS ABABA UNIVERSITY.
Zhang, S., Yang, K., Li, M., Ma, Y., & Sun, M. (2018). Combinational Biophysical Composition Index (CBCI) for effective mapping biophysical composition in Urban Areas. IEEE Access, 6, 41224–41237. https://doi.org/10.1109/ACCESS.2018.2857405
Zhou, W., Huang, G., Troy, A., & Cadenasso, M. L. (2009). Object-based land cover classification of shaded areas in high spatial resolution imagery of urban areas: A comparison study. Remote Sensing of Environment, 113(8), 1769–1777.

No competing interests reported.

Supplementary.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Detection of scattered settlement in and around Babile Elephant Sanctuary and threats it pose using robust remote sensing method, Eastern Ethiopia.

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and Methods

2.1. Study Area

2.2. Data source

2.3. Image Preprocessing and interpretation

2.3.1. Cloud masking

2.3.2. Feature selection and training data preparation

2.3.3. Classification method and accuracy assessment

2.3.4. Settlement map integration method

2.4. Spatial analysis of settlement distribution

3. Results and Discussion

3.1. Settlement Mapping

3.2. Comparison of the produced map with an existing dataset

3.3. Accuracy assessment

3.4. Spatial Analysis of Settlement Distribution

3.4.1. Settlement clusters

3.4.2. Settlement distribution by proximity

3.4.3. Non-settlement area

4. Conclusions

References

Additional Declarations

Supplementary Files

Status:

Version 1