Research Forest Loss and Susceptible Area Prediction at Sefwi Wiawso District (SWD), Ghana

Forests provide immeasurable merits for the economies of most developing countries. Forests in developing countries experience harmful human-induced impacts such as unregulated removal of biodiversity and unsustainable land conversion. The Sefwi Wiawso District (SWD) located in Ghana, which includes portions of six protected forest reserves (FRs) such as Muro, Tano Suhien, Tano Suraw, Suhuma, Sui River, and Krokosua, is the subject of this study. The impacts of selected spatial variables on forest losses were examined using retrospective and predictive approaches. Past deforestation patterns were analyzed using classiﬁed Landsat 5 and 7 imagery from 1984 to 2017. Pixel areas in hectares (ha) from land use land cover (LULC) classiﬁcations were used to detect land cover classes that were vulnerable to potential loss. The study also carried out a simple forest prediction using the simple moving averages (SMA) forecasting model based on the past and present deforestation patterns from LULC classiﬁcation. The results showed that 3587.49 hectares (ha) of protected forest cover was converted into agricultural lands and barelands. In addition, 2532.96 hectares (ha) was converted from close forest to nonforest land cover from 2000 to 2017, which is equivalent to a 16% reduction in close forest cover within the FRs in the SWD. This loss was also 11% higher than close forest areas between 2000 and 2010. SMA forecasting showed that from 2017 to 2024, 877.38 hectares (ha) of close forest resources will convert to open forest resources and other nonforest land cover. Subtle accessibility routes such as navigable rivers and unoﬃcial roads are the key instigators of protected forest clearance in the Sefwi Wiawso Forest District (SWFD). The SWFD is surrounded by many communities and is susceptible to uncontrollable biodiversity removal due to lack of proper monitoring of agricultural practices, mining operations, fuelwood collection, and illegal hunting, which represents a means of livelihood for the forest fringe community dwellers. The research serves as a benchmark for similar studies in eﬀorts to investigate, measure, and project land cover change in protected forest areas.


Introduction
e mechanics involved in understanding the depletion of forests have always been problematic. In measuring the rate of forest loss, past research studies have relied on indicators such as location, forest type, climate conditions, rainfall patterns, proximate forest dependencies, infrastructural development, population growth, and livelihoods [1,2]. Industrial wood, fiber, food, medicines, and firewood are few benefits fringe communities derived from the protected forest. However, unsustainable forest removal for short term communal gains has resulted in adverse effects such as biodiversity loss, unpredictable climate conditions, and destruction of terrestrial biodiversity ecosystems [3,4].
Instigators of forest loss are events, structures, or practices which directly or indirectly cause the conversion of forest land cover types to nonforest land cover types [5]. e identification of direct and underlying instigators of forest loss is essential for developing plans and schemes required for sustainable forest resource management [6,7]. Analyzing the causes and dynamics of deforestation leads to better results at the local and rural levels, since the data collected are small in scale and thus indicate the root causes [8]. International conservation organizations have committed at least $3.4 billion to help mitigate deforestation in Africa since 2000, but the effects of these attempts are not readily evident due to short term objectives and goals [9].
Ghana is one of the several developing countries encouraged to reduce forest losses for reducing emissions from deforestation and forest degradation, conservation, and enhancement of carbon stocks (REDD+) benefits [10]. e development of national strategies and action plans for REDD+ in 46 developing countries has shown that commercial and subsistence agriculture is the main instigator of forest loss, whilst unsustainable logging is identified as another major degradation enhancer [11]. Human-induced factors such as fuelwood processing, charcoal production, wildfires, and livestock grazing have been identified as secondary instigators of forest removal in developing countries [11]. Despite forest management interventions by the Forestry Commission (FC) of Ghana, there is a persistent decline in protected forest infrastructure within the SWFD. Continuous removal of forest biodiversity without sufficient replacement has led to effects such as desertification, floods, deforestation, disease prevalence, and pest infestations [12]. High forests covered 79,511 km 2 of land area in Ghana in 2000 but stood at 7,951.12 km 2 in 2016 [13].
e development of the geographic information system (GIS) and remote sensing (RS) techniques has proved to be important in addressing challenges encountered when quantifying and identifying forest loss instigators [14][15][16][17][18]. Sustainable conservation methods use GIS/RS techniques to try and remain ahead of the issue [14,15,19]. Remote sensing analysis of time-series satellite imagery has helped to classify current and past forest cover patterns and further identified trends of forest conversion from which future forecasts can be made [20][21][22]. It is also easier to control forest biodiversity when ground forest data are regularly collected in the GIS environment [17,23]. e advent of unmanned aerial vehicles (UAVs) enabled the timely retrieval and study of forest spatial data without too much exertion [17,23]. e International Union for Conservation of Nature ranks protected areas within SWD as category 1a where limitations on accessibility are placed on proximate dwellers, and forests are reserved for sustainable management, scientific, and biological research. However, the FRs within the SWD do not reflect these strict protective guidelines. It is based on this that the study investigates, quantifies, and forecasts forest land cover change for protected forest biodiversity in SWD using accessible roads, agriculture, and community density as spatial indicators. e study first specifically analyses the risk levels of protected forest areas based on proximity of access roads within the SWFD. e study goes on to assess the impacts of agricultural expansion and community location density on protected forest land cover conversions in SWFD. e study then predicts forest loss dynamics in the near future based on the aforementioned drivers.

Materials and Methods
e study used several time-series satellite imagery, road, and community shapefiles within the SWD to reveal conversions from forest cover to other land cover based on selected drivers within the FRs [16]. Forest cover images were classified into close forest, open forest, farmland, grassland, and bare land cover categories using the Gaussian maximum likelihood algorithm. Accuracy assessment was carried out to examine the accuracy of the classifier's prediction. Road and community shapefile data were processed using the ArcGIS toolbox to define road buffer distances and community location intensity. Resulting road and community datasets were integrated with results from LULC classification to reveal forest changes and land cover conversions in hectares (ha) within the SWD based on the drivers identified.

Study Area.
e SWD (Figure 1) which has an area of 101,060 hectares (ha) is located in the Western North region of Ghana and contains the portions (Table 1) of the forest reserves (FRs) used for the study. It is bounded by Asunafo South, Atwima Mponua, Sefwi Bibiani-Anhwiaso, Wassa Amenfi West, Sefwi Ankontombra, Bodi, and Juabeso political districts. It falls within the moist semideciduous ecological zone. e FRs inside the study area are colonized by dominant timber tree species such as Heretiera utilis, Cynometra ananta, and Celtis milbraedii, which have several medicinal and research purposes [24]. e FRs also provide natural habitats to mammals including the giant rat, the mona monkey, the red river hog, and Maxwell's duiker to name a few [24]. e inhabitants of the SWD have benefited from the timber utilization contracts (TUC) policy which include the development of schools, health facilities, and improved road networks. is has consequently improved income generation and increased the level of employment in the district [25].
A recent study in the southern part of Ahafo and the northern part of the Western North regions which contain the study area depicted an increasing rate of deforestation [13]. Primary and secondary forest loss had an annual rate of 1.9% over 25 years (1986-2011), but currently has an annual rate of 2.3% [13].

Landsat Imagery Preprocessing and Classification.
e Landsat images used were from 1984, 1990, 1996, 2000, 2003, 2010, and 2017. Raw landsat imagery was acquired from the United States Geological Survey's (USGS) Earth Explorer using the SWD's area of interest (AOI). Instances of haze were removed through earth observation data repository. Histogram equalization was carried out in ArcGIS. Ground reference points representative of five land cover categories (Table 2), namely, close and open forest, farmland, grassland, and barelands used were collected. Band combinations 2 International Journal of Forestry Research (Table 3) for false color composites in Landsat 5 and Landsat 7 images were used to identify various land cover within the study area for better classification results [27]. e digital numbers (DNs) of the acquired imagery were subsequently sampled and trained using the Gaussian maximum likelihood classification algorithm (Figures 2  and 3).
is algorithm was preferred because of the Bayes theorem decision-making, which accounts for some forest classes occurring in higher or lower pixel quantities than the average [28].

Accuracy Assessment.
e study assessed the classifier's ability to accurately and quantitatively identify how efficient pixels were sampled into their correct land cover categories. A total of 350 geographically verified pixels (ground truths) were created on the images. ese training samples were randomly assigned to each class and were proportional to the land cover sizes. Error matrix report tables for all LULC classifications (Tables 4 and 5) showed the relationships between ground truth data and the corresponding classified data [27]. Statistical accuracy metrics (Tables 6 and 7) which   International Journal of Forestry Research    [27,31]. e columns (Table 4) show the classes in the validation (ground truth) pixel set, and the rows show the classified pixel sets. e overall accuracy which is the percentage of correctly classified points from the total number of points is 79% with a kappa of 0.738.

Analysis Using Forest Loss Instigators
2.4.1. Road Proximity. Road shapefiles within the SWD were acquired and processed into three buffer distances. On-field verification coordinates and Google Earth Imagery were used to verify the current existence of road segments, junctions, and turns used in this study. e road segments used had a total distance of 271.20 kilometers. As a forest loss indicator, variable accessible road segments within FR areas raised red flags as they were suspected to expose the FRs to different forms of human interventions and deforestation risks [33]. e three road buffers set were based on proximity to FRs. High-, medium-, and low-risks identities were given to the road buffers based on the distance to FRs [33,34]. Referring to [34], this study devised a minimum distance interval (Table 8) for road buffers that would trigger early detection of forest loss in FRs. A workflow of road shapefile processing and subsequent LULC integration is shown in Figure 4.     63  3  2  0  5  73  63  2  Open forest  10  87  8  0  4  109  87  3  Farmland  5  3  64  2  7  81  64  4  Grassland  4  5  0  47  0  56  47  5  Bareland  3  5  4  1  18  31  18  Total  85  103  78  50  34 350 279     shapefiles for the SWD were acquired and used for the analysis. Eight hundred and ninety-six forest fringe communities were spatially verified and found to be within the SWD. A density raster created from the community point shapefiles was used for spatial autocorrelation [35]. Communities whose boundaries coincided with each other were integrated through a specified XY tolerance [35]. e point density radius of the communities within the SWD was computed through repeated spatial autocorrelation. Random distance values were compared to the corresponding zscores, and a line graph was created to measure the peak zscore [35]. e z-score recorded the intensity of community clustering. A community density raster ( Figure 6) was calculated and integrated into the LULC dataset ( Figure 7).

Simple Moving Averages the LULC Forecasting Model.
Predicting the pixel area conversions for 2020 and 2024 LULC was achieved with the simple moving averages (SMA) forecasting model [36]. e model calculates a current image average by adding recent changes and dividing by the number of periods in the calculation average [36]. e mean absolute deviation (MAD) was calculated for each average between two selected years. e cell information (Figure 8) of the image was extracted and counted for each class. e pixel area (PA) per class was determined by multiplying the width of image resolution by the cells count per class. Pixel areas for each category were extracted after average prediction and used for the arithmetic. e pixel area (PA) of the classified imagery for the years 1990, 1995, 1996, 2000, 2003, 2010, and 2017 were used to predict 2020 and 2024 LULC imagery. e predicted land cover year was calculated by averaging the previous two years separated by the same interval if major LULC changes were expected within that timestamp [1]. An example is shown in equation (1) for predicting the pixel area for 2010 LULC. e error between a predicted year and  International Journal of Forestry Research the original land cover is seen in equation (2) and catered for by the MAD calculated in equation (3). LULC for 2020 and 2024 is calculated in equations (4) and (5).
where PA is the pixel area of the classified image, PPA is the predicted pixel area of the classified image, Err is the error between the original and predicted forest land cover pixel areas, and MAD is the mean absolute deviation. e forecasting method was based on the assumption that past trends in forest loss will continue in the future.
(i) Cell counts per class used in calculating the pixel area per class and performing pixel arithmetic between categorical LULC classes (ii) Two time stamps used in calculating the predicted pixel area for equal interval time points Community clusters in SWD Figure 6: A community density raster created through spatial autocorrelation.
(iii) (Shown yellow) Predicted pixel area and original pixel area for a time interval used to calculate the mean absolute deviation and predicted pixel area error (iv) Predicted pixel area for 2024 using the predicted pixel area error and mean absolute deviation

Road Proximity Impact.
After the integration of LULC with road buffer datasets, the risk status of the FRs was visualized (Figure 9 and   (Table 11). e relatively low area conversion was due to intensive and enhanced community involvement in sustainable forest management particularly in Krokosua FR [37]. Open forest cover area of 141.39 hectares (ha) was converted to farmland cover from 2000 to 2010. Open forest cover area of 1313.28 hectares (ha) was converted to close forest cover, and 2321.46 hectares (ha) of close forest cover

Risk level Forest reserve
High  were seen to be within the accessible zones of clustered fringe communities within the SWD. Tano Suhien, Tano Suraw, and Sui River FRs were inside the range of immediate community coverage ( Figure 6 and Table 13). Based on closeness to highly dense community locations, eight risk classes of disturbance were created and tabulated to visualize threat to forest cover removal posed by fringe communities. It was revealed that 815 hectares (ha) of forest cover was found to be between high and critical disturbance risks (Table 14), since they were in direct contact with fringe community dwellers. Also, 4622.94 hectares (ha) of close forest cover within the high disturbance levels were removed from 2010 to 2017 with an increase of 3036 hectares (ha) in farmland cover in the same period.

Prediction
Results. e simple moving averages forecasting method revealed that from 2017 to 2020, 144.31 hectares (ha) of close forest cover will be removed together with 125.44 hectares (ha) of open forest cover (Table 15). However, there will be an increase of 540.63 hectares (ha) in open forest cover from 2020 to 2024. Farmland cover will increase by 71.10 hectares (ha) from 2017 to 2020 and increase by 140.13 hectares (ha) from 2020 to 2024.

Discussion
e pseudo images from LULC classification were validated using the error matrices. e subsamples for training and testing were systematic and was influenced by the report of [38] on the fitness of different ground truth sampling plans in satellite image accuracy assessment. e adapted assessment technique, however, allows room for improvement as the overall accuracy is biased since the number of sampling points per class is not the same. e close forest cover areas in the images were reduced by 3210.022 hectares (ha) in four of the six FRs within the road buffers created. Farmland cover was not dominant in the year 2000 but had increased drastically by the year 2017.
is indicates that FRs nearest to accessible roads are targeted for illegal entry and economic activity. FRs whose boundaries are outside the 3 km road buffer zone show the lowest forest cover conversions. e road buffer intervals adopted here triggers early forest loss detection in protected forest areas relating to the study conducted by [33] in Amazon, which identified that unofficial roads within 5.5 kilometers of forest areas have bearings on deforestation rates. From the findings in this study, accessible roads near FRs were observed to provide gateways into FRs for illegal and destructive activities by forest fringe community dwellers.
Observations from findings also reveal small instances of farmland cover in the year 2000. Subsequently, farmland cover in the six FRs increased steadily from 2000 to 2010 but sharply from 2010 to 2017. Priority was therefore given to large scale agriculture over forest conservation, since it is the primary source of livelihood amongst the forest fringe communities. e farmland cover conversions revealed by this research highlight a missing stance in early research  studies like [39] which gives a general view that human activities such as agricultural expansion and illegal and unsustainable logging are responsible for the degradation of 85% of Ghana forest areas. Farmers tend to replace forest biodiversity with agricultural lands, which are considered more financially beneficial within the short term. Cash crops such as cocoa, cashew, and rubber are as a result more important to farming communities than biodiversity conservation. e three most disturbed FRs based on community influence suffered a drastic reduction in the close forest area and an increase in the farmland area. In situ data collection and interviews revealed that forest fringe settlers engage in high-intensity small scale mining, large scale farming, chain saw operations, and illegal hunting within the nearby reserves [40]. FRs close to densely populated community locations showed rapid conversions from forest to nonforest cover.
Close forest areas will reduce steadily, while open forest areas will increase sharply from 2020 to 2024 after a steady reduction from 2017 to 2020. e simple forecasting method looks smooth especially over the selected period. Barring any forest replacement interventions, farmlands will continue to increase.

Conclusions and Recommendations
Destructive human interference is promoted by accessibility through close accessible roads. Easy access routes enable proximate dwellers to easily access FRs undetected to illegally remove biodiversity. is research was unable to take into account the considerable number of unofficial routes through which proximate dwellers can access protected forest areas.
Conversion from forest to farmland cover is at a very high rate. Agricultural expansion rates continue to increase within FRs. e IUCN's categorization of FRs is undermined by the intrusion of extensive farming areas. e severe damage in FRs as a result of uncontrolled daily agricultural expansion could be monitored in further studies.
Population and community growth increase the intrusive risks in protected areas. Illegal human activities such as mining, farming, logging, and reckless fuelwood collection are practiced by fringe communities as a means of sustenance.
Year 2020 and 2024 predicted imagery showed trends of forest removal and farmland increase in the future. If conservation interventions do not improve, the adverse community impacts on FRs will be irreversible and unsustainable.
e stakeholders of the forest should consider real-time monitoring procedures, which account for subtle modes such as navigable rivers and skid trails areas which allow easy access to FRs. e emergence of illegal farms within protected forest areas should be discouraged. Laws concerning illegal entry into protected areas must be better enforced.     16 International Journal of Forestry Research e focus should increase on community-based participation in forest protection. Short interval imagery should be considered in the analysis of this nature to better depict unwanted conversions at an early stage.
is will help identify impacts of dominant nonforest land cover, which is responsible for the conversion, and also measures the rate at which conversion has taken place with confidence.

Data Availability
e data used to support this study are included within this article.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper.