Landsat MSS (20/08/1975), Landsat TM (30/09/1995) and Landsat ETM+ (20/07/2012) imagery with better visualization (less than 15% cloud cover) from the Global Landcover Facility (http://www.landcover.org) were selected for the study. Datum was set to WGS 84 and referenced to Universal Transverse Mercator (UTM) zone 37 South. All images were orthorectified using ground validation points, Digital Elevation Model (DEM) and aerial photos as a reference. Landsat images were resampled to a common resolution pixel (30 × 30 m) using bilinear resampling to ensure consistency in all image scenes. First order polynomial transformation was applied at image registration to correct for any shifts. It was deemed necessary to simulate atmospheric interactions between the sun and sensor pathways for the imagery used. Therefore, a radiative transfer model in Atmospheric and Topographic Correction (ATCOR) module in Erdas Imagine 2013 was used for atmospheric correction. ATCOR masks haze, cloud, water and enhances pixel visibility. In this study, we used the MODerate resolution atmospheric TRANsmission (MODTRAN) code to retrieve the atmospheric parameters for ATCOR from the look-up table as ground-based reflectance and atmospheric data were unavailable. Digital number values were then converted to reflectance based on metadata provided with the Landsat images (Chander et al. 2009, Guanter et al. 2009, ERDAS and Geosystems 2011).

A supervised maximum likelihood classifier was adopted for classification (Liu et al. 2002, Manandhar et al. 2009, Tseng et al. 2008, Xi 2007). The technique is based on statistical probability that assigns pixel values to the category with the highest likelihood (Aldrich 1997, Dean and Smith 2003, Ince 1987). Spectral signatures were created and applied in categorizing similar pixels in the entire image using eight polygons representing training data sets for each habitat class. A color composite of 3, 4 and 5 bands were used to facilitate visual interpretation while the Gaussian distribution function was applied in the stretching process. The image was classified into four class categories namely: Woodland, Grassland, Dense forest and Less dense forest. A total of 82 field ground data points, archival high resolution aerial photographs, interviews and expert opinion were used to validate the classified images. Confusion matrices were then created to compare reference data with the maximum likelihood prediction and for calculation of the overall accuracy (OA), producer’s accuracy (PA) and user’s accuracies (UA). Overall accuracy is a percentage (%) between correctly classified classes and the total number of test reference data, while producer’s accuracy is the probability of a specific class being correctly classified. User Accuracy is the possibility that a sample of a specific class represents the category on the ground.

Fragstats metrics were extracted from all processed Landsat images. Fragstat metrics offer a distinct capacity to determine a landscape’s spatial configuration, hence valuable in understanding landscape change arising from fragmentation (Cushman 2006, Jorge and Garcia 1997, Saikia et al. 2013, Millington et al. 2003). All classified images were converted to ASCII format in ArcGIS 10.2. A C-program, a raster version inbuilt within Fragstas that accepts ASCII image files was applied using the eight cell rule. The ASCII format scenes were imported into Fragstats and ASCII built-in-algorithm selected for running the Fragstats model. Three multi-level structure metrics were selected at patch, class and landscape level (McGarigal and Cushman 2002). Metrics relevant in explaining the magnitude and extent of fragmentation were then selected from the 1975, 1995 and 2012 image scenes. A total of 155 samples were randomly selected and extracted. As recommended by McGarigal and Cushman (2002) two metrics i.e. perimeter area relationship and patch area were statistically used to test the magnitude of fragmentation. Mann-Whitney U and Post hoc ANOVA tests were used to evaluate differences among patch areas in all the years. The Games-Howell was used to determine forest fragmentation. The indices used in this study are briefly described in Table 1.

The overall accuracy for 1975, 1995 and 2012 image scenes was 78.26%, 84% and 76.54% respectively (Table 2). Changes in total area coverage were observed in all years (Figures 2a, b and c).

Figure 2.

Land use land cover (LULC) maps in 1975 (a), 1995 (b) and 2012 (c).

The study findings showed substantial land modification in most of the cover types during the study period i.e. decline in dense forest (31, 675.70 hectares) and less dense forest (by 11, 267.38 hectares) and increase in grassland (21, 230.01 hectares). However, changes in areas covered by woodland were inconsistent, i.e. increase by 15,884.46 hectares between 1975 and 1985 and decline by 8, 182.03 between 1985 and 2012) – Figure 2.

There was a transformation in habitat extents within the study area. Significant losses were recorded for dense forest (31, 675.70 hectares) and less dense forest (by 11, 267.38 hectares), however, there was a steady increase in areas covered by grassland. Based on field study observations, these changes can be attributed to expanding agricultural fields and increased exploitation of timber and non-timber products to meet the increasing urbanization demand in Morogoro district. This finding is in agreement with Burgess et al. (2001) and Burgess et al. (2002) who found a substantial decline of dense forest in the Uluguru mountains due to urbanization and agricultural (Burgess et al. 2002, Burgess et al. 2001). In other parts of Tanzania, related studies established effects of reduced tree density to land modification (Yanda and Shishira 1999, Munishi et al. 2010). Habitat modification could also be attributed to general population increase in non-urbanized areas, also known to influence its spatial configuration (Fischer and Lindenmayer 2007).

As aforementioned, there was a general decrease in area covered by dense and less dense forest habitat. A decreasing trend in the extent of total habitat coverage relates to deleterious fragmentation as effects of habitat fragmentation are dependent on habitat size (Fahrig 2003). Furthermore, perimeter-area results in this study show distinct differences in woodland and grassland habitats. In most instances, high perimeter-area relationship characterizes rapid rate of fragmentation underlying the two landforms e.g. Jha et al. (2005) and McGarigal (2006). Woodland habitat displays a patchy type of deforestation, shown by an increased patch number between 1975 and 2012. The slight decline in patch number can be attributed to the strong traditional leadership forest maintance authority in the 1970s, a responsibility that has now been taken over by the Tanzanian Government that permits logging and farm allocations. Dynamics in mean patch area were observed in the woodland and less dense forest. Notable was the gradual decrease in patch size, while patch number increased by 412 and 391 in dense forest and woodland respectively, an indication of fragmentation patterns in the area earlier observed by Jha et al. (2005). On the other hand, patch area was ideal in characterizing distinct areas with analogous environmental conditions, where patch boundaries are distinguished by discontinuities in environmental character states relevant to the organism or ecological phenomenon under consideration. A combination of patch density (PD), PARA and mean nearest neighbor distance are considered profound in estimation of the extent of fragmentation in each of the habitats analyzed (Jha et al. 2005). Patch density and PARA are regarded as important in fragmentation assessments, particularly in natural ecosystems because they have a strong influence on ecosystem functioning and ecological processes (McGarigal 2006).

Similarly, a distinct variation in patch number was observed. Woodland and less dense forest had the highest patch number across the years. This can be attributed to the great extent of fragmentation resulting from natural resource exploitation. Furthermore, their vicinity to Morogoro town and management by local authorities may be possible drivers increasing their susceptibility to fragmentation (Fahrig 2001, Fahrig 2003, Wiens 1995, McGarigal 2006, Fischer and Lindenmayer 2007). The woodland habitat had a relatively greater patch density, signifying higher spatial heterogeneity. In addition, the largest patch index was associated with less dense forest while least values were associated with the grassland habitat. This provided information on least and most fragmented landscapes, a good indicator of minimum area requirements for species survival (McGarigal 2006). In addition, the largest patch index, another good indicator for species survival was significant in the less dense forest compared to the rest of the habitats (Rutledge 2003).

Dense forest and woodland had the greater edge density. This could be attributed to increased exposure to farmlands and settlements prevalent in the area. Edge effects characterize the biophysical state of ecosystems at the periphery or in the neighborhood. This is because increased habitat fragmentation exposes habitat to edge effects, compromising the ability of an ecosystem to provide relevant goods and services (Murcia 1995). This limits a habitat’s long-term ability to sustain a population as it intensifies species mortality rate (Fahrig 2003). It also influences occurrence of native species populations (Murcia 1995) and ensures that the interaction of species in disturbed environments remains restricted, advancing their mortality risk (Rutledge 2003). Related literature also found a high intensity of fragmentation associated with more edge effects through exposure of contiguous habitats to solar radiation and soil moisture to drier heat conditions (Rutledge 2003).

Games-Howell test results showed a significant level in the perimeter area relationship (p≤0.05). This could be explained by the fact that less dense forest adjoins dense forest, taking up regions dominated by woodland. It is also possible that the on-going fragmentation is a major driver of conversion of dense forest and woodland to less dense forest. Potential socio-economic drivers could be a result of the expanding Morogoro town and increasing agricultural fields in the adjacent local regions. Similarly, other studies showed how adjoining activities influence intact habitat ecosystems as a result of their structural configuration (Echeverría et al. 2007).

Anthropogenic activities significantly influence habitat fragmentation in the region. For instance, extensive farming and urban growth are possible drivers to habitat modification and fragmentation. The area has a conducive montane climate that supports subsistence farming, a prevalent socio-economic activity in the region (Burgess et al. 2007, Swetnam et al. 2011, Yanda and Shishira 1999). This seems to significantly influence all the four habitats. Increasing population growth and consequent increase in settlement and farmlands may have extirpated important fauna and flora in the Ulugurus (Bjørndalen 1992, Burgess et al. 2002, Burgess et al. 2001, Burgess et al. 2007, Swetnam et al. 2011, Hall 2009, Yanda and Shishira 1999) – Figure 5. Habitat fragmentation in the study area can also be attributed to a complex nexus of socio-economic processes (Kessy et al. 2016, Rosales 2008). These processes act at various scales i.e. international (global forest products market growth, commercialization and urbanization), national (changing population, growing local markets and national legislation and governance) and local conditions (livelihoods and levels of poverty) (Wehkamp et al. 2015, FAO 2007, Daly and Farley 2004, Czech 2013). Kessy et al. (2016) for instance notes that local and international demand for timber and agricultural commodities in a globalizing world are major drivers to forest fragmentation in the area. Globalization, with its characteristic scramble by the developing countries to increase their market share on the global marketplace has increased pressure on existing forests and forest land (Hecht and Saatchi 2007, Rosales 2008).

Figure 5.

Drivers to fragmentation, note the settlements in the valley and cleared forest in the background and fore ground for crop farming and grazing, respectively (A) and small scale maize and banana fields within the forest in (B).

To forestall some of the problems earlier highlighted, the study area, identified as biodiversity hotspots with important ecological functions such as groundwater recharge, surface flow and animal habitat need to be protected from the impacts of land modification and fragmentation. Implications of habitat modification and fragmentation in Morogoro region can be better deciphered through the impact on habitat structure and species losses. The increased habitat losses, mainly attributed to anthropogenic factors may negatively influence genetic diversity and lead to losses of potentially useful genes originally accommodated in intact areas (Ojoyi et al. 2015, Burgess et al. 2007, Swetnam et al. 2011, Hall 2009, Yanda and Shishira 1999, Shirima et al. 2011). Therefore, we recommend that mitigation measures should be adopted to ensure protection and management of these fragmenting habitat ecosystems. To optimize mitigation measures, the adverse effects of habitat modification and fragmentation need to be understood by all stakeholders. In addition, policy measures and sustainable bottom-up approaches to management and conservation of forest resources should be instituted in the region.