Appropriate satellite imagery acquisition was done with highly consideration of cloud cover, the seasonality and phonological effects (Kashaigili et al., 2006). Clouds free satellite images with the interval not less than ten years from 1990 to 2015 (Table 1) sourced from USGS-GLOVIS (https://glovis.usgs.gov/) were used in assessing temporal and spatial variation of land use/cover change in the study area.

To ensure accurate identification of temporal changes and geometric compatibility with other sources of information, images were pre-processed whereby geo-correction was conducted to rectify precisely matching of images. Band stacking and image enhancement was performed using different color composite band combinations to reinforce the tonal distinctions. Images were registered to the

Universal Transverse Mercator (UTM) map coordinate system, Zone 36 South, Datum Arc 1960 (Arc 1960/UTM zone 36S). Image mosaic was conducted to merge together images of the same year with same path and different row so as to create a single image that covers the entire catchment (Figure 2).

Supervised image classification using Maximum Likelihood Classifier (MLC) was conducted to classify the images. Maximum likelihood classifier is one of the most popular classification algorithms which calculates posterior probability of a pixel belonging into corresponding class based on Bayes theorem (Congedo, 2016). The algorithm used probability density functions to assigns the pixel to the most likely (highest probability value) class (Lillesand & Kiefer, 2002) which increase the accuracy of classification output to resemble to the real world. A base-map for year 2015 was created and used for ground truthing fieldwork. The basis of land cover type nomenclature adopted “The National Forest Resources Monitoring and Assessment (NAFORMA) Field Manual Biophysical survey of 2010” with few modifications to reflect the actual ground conditions (Table 2).

The training sites for MLC were identified by inspecting an enhanced color composite imagery. Areas with similar spectral characteristics were trained and assigned respective classes. The Semi-automatic Classification Plugin (SCP) available in QGIS 2.12.1 was applied to classify the image into twelve distinct land cover classes.

Ground truthing was conducted using a hand-held Global Positioning System (GPS) with a base-map of year 2015. The essence of conducting ground truthing was to verify different covers types as described on the base maps and for collection of ground points for the classification accuracy assessment. The error matrices (Congalton, 1991) were prepared and used in computation of Kappa coefficient (K) for the classification accuracy assessment of final image classification. The probability of a classified land cover map corresponds accurately with the ground truthing data assessed by the user’s accuracy (Jansen, 2004) and the measures of the agreement of classified maps and ground truthing data assessed by Kappa statistics (Lillesand et al., 2014).

K = ∑ i = 1 r x i i − ∑ i = 1 r ( x i + × x + i ) N 2 − ∑ i = 1 r ( x i + × x + i ) (1)

where N is the total number of sites in the matrix, r is the number of rows in the matrix, x_ii is the number in row i and column i, x_+i is the total for row i, and x_i+ is the total for column.

To analyses the changes between different time epochs, change detection analysis was performed. Post classification comparison was used to quantify the extent of land cover changes for the periods 1990, 2005 and 2015. The estimation for the rate of change for the different land covers was computed based on (Kashaigili & Majaliwa, 2010).

% Cover change = Area i year x − Area i y ear x + 1 ∑ i = 1 n Area i year x × 100 (2)

Annual rate of change = Area i year x − Area i year x + 1 t years (3)

% Annual rate of change = Area i year x − Area i year x + 1 Area i year x × t years × 100 (4)

Area i year x is the area in hectares (ha) of cover i at the first date;

Area i year x + 1 is the area in hectares (ha) of cover i at the second date;

∑ i = 1 n Area i year x is the total cover area at the first;

t_years is the period in years between the first and second scene acquisition dates.

Markov chain is a statistical tool that describes the probability of land use to change from one time period to another by developing a transitional probability matrix between first period and second period based on the spatial neighborhood effects (Wang et al., 2004; Al-Bakri et al., 2013; Araya & Cabral, 2010). A spatial neighborhood effect is the state of neighboring cells to influence the transition of a given cell into different states. This model was based on using and evaluating land use layers of previous years to predicting the spatial distribution of land uses in the future (Wu & Silva, 2010). For better simulation of temporal and spatial patterns of land use changes in quantity and space, the combination of two techniques Markov chain analysis and Cellular automata (CA-Markov) were used.

The simulated model was developed by using IDRISI Selva v.17.0 software (Rutherford et al., 2015) and it involved two main stages which are calculating conversion probability (conversion probability matrix, conversion area matrix and layers of conditional probability) done by using Markov chain analysis, and the second stage was spatial specification of land use coverage simulated based on Cellular Automata spatial operator and multi criteria evaluation (MCE).

In the developing CA Markov model, the classified land use map of 1990 which represent past, and 2015 which represent present time developed in QGIS 2.12.1 were converted into IDRISI data format and selected to be input data into the model, to calculate matrices of conversion probabilities and conversion areas (Transition area matrix and transition probability matrix).

The transition probability matrix (Table 3) expresses the likelihood (probability) that a pixel of a given class that will change to any other class (or stay the same) in the next time period. The transition areas matrix (Table 4) expresses the total area (in cells) expected to change from the year 2015 to the year of 2040 according to those changes happened from 1990 to 2015.

For model validation the simulated land use/cover map for 2015 was compared with the actual satellite derived land use/cover map based on the Kappa statistics. Then, standard Kappa index was used to check whether the model is valid or not (usually the Kappa Index for a valid model is >70%) (Wen, 2008). If the model has the Kappa Index less than 70% then the suitability map for the land covers and filter used should be repeated based on several considerations. VALIDATE tool was used to compute Kappa statistics for the projected land use/cover.

The classified maps showed good agreement with the real world as indicated by

FR: Forest, PL: Plantation, RF: Riverine forest, WTR: Water, WET: Wetland, WD: Woodland, GR: Grassland, WR: Wooded rock, CW: Cultivated woodland, BS: Bushl, CLT: Cultivated land, BLT: Built up.

FR: Forest, PL: Plantation, RF: Riverine forest, WTR: Water, WET: Wetland, WD: Woodland, GR: Grassland, WR: Wooded rock, CW: Cultivated woodland, BS: Bushland, CLT: Cultivated land, BLT: Built up.

overall classification accuracies of 99.79%, 98.43%, and 99.25% respectively, for 1990, 2005, and 2015 with their corresponding Kappa statistics of 0.99, 0.98 and 0.99 respectively (Table 5). It is recommended that Kappa value for a good classification performance should be more than 0.80 (Jensen, 2005; Lillesand et al., 2004) and the minimum level of promoted classification overall accuracy in identification of land use classes should be at least 85% (Anderson, 1976).

The land use and land cover maps for the year 1990, 2005 and 2015 are presented in Figure 3. Generally, maps show variation in land use and land cover coverage between time periods. Table 6 presents the coverage of each land use and land cover class between 1990 and 2005 and between 2005 and 2015. The extent of LULCC including the area change, percentage area change, annual rate of change and percentage annual rate of change for the Little Ruaha River Catchment are summarized. The increased and decreased amount is represented by positive (+) and negative (−) signs respectively.

The results indicate that for the period between 1990 and 2005 the area under Forest which occupied 39,872 ha (6.26%) in 1990, decreased to 22,957 ha (3.6%), indicating a decrease of about −42.2% from its original coverage. Likewise, riverine forest and woodland decreased from 5878 ha (0.92%) and 109,692 ha (17.22%) to 2746 ha (0.43%) and 72,809 ha (11.43%) respectively, showing a decrease of −53.28% and −33.62% for riverine forest and woodland respectively. Water and wetland areas declined from 1752 ha (0.28%) and 19,157 ha (3.01%) to 1202 ha (0.19%) and 11,785 ha (1.85%) indicating a loss of −31.39% for water and −38.48% for wetland between 1990 and 2005. At the same time, cultivated land and built up area showed an increase from 106,782 ha (16.76%) in 1990 to 109,047 ha (17.12%) in 2005 and from 9408 ha (1.48%) to 11,674 ha (1.83%) respectively, indicating a gain of about +2.12% and +24.09% respectively.

For the period between 2005 and 2015, the Forest, riverine forest, woodland, water and wetland declined to 15,950 ha (2.50%), 1083 ha (0.17%), 58,554 ha (9.19%), 757 ha (0.1%) and 5622 ha (0.88%) respectively, indicating percentage loss of −30.52%, −60.56%, −19.58%, 45.34%, and −52.30% respectively. For the same period of time, cultivated land and built up area increased to 143,468 ha (22.52%) and 13,765 ha (2.16%) respectively, indicating a gain of +31.57% for cultivated land and +17.91% for built up area. The area under Forest decreased at a rate of −1127.67 ha/year (−2.83%/year) over a period of 15 year (1990-2005), and −700.70 ha/year (−3.05%/year) over a period of 10 years (2005-2015), likewise, riverine forest, woodland showed a similar trend of decline for both time periods.

The overall land use land cover change from 1990 to 2015 indicate that forest, riverine forest, water, wetland and woodland has declined by 60%, 82%, 63%, 71%, and 47% respectively from their original coverage, while plantation, grassland, bushland, cultivated land and built up area increased by 18%, 25%, 44%, 34% and 46% respectively. Result shows that woodland, forest, and wetland are the most altered ecosystem where by annually decline at the rate of 2045.52

FR: Forest, PL: Plantation, RF: Riverine forest, WTR: Water, WET: Wetland, WD: Woodland, WR: Wooded rock, CLT: Cultivated land, CW: Cultivated woodland, GR: Grassland, BS: Bushland, BLT: Built up area.

ha/year, 956.88 ha/year and 541.40 ha/year respectively. Riverine forest decline at the rate of 191.80 ha/year while water bodies decline by 43.80 ha/year.

The overall gain and loss and the net change of each land use and land cover category between the period 1990-2005, 2005-2015 and 1990-2015 are summarised in Table 7, and change detection matrix showing the transformation of each land use category presented in Tables 8-10.

The results, clearly reflecting the dynamics of land cover classes in the study area. Forest, Woodland and Wetland were found to be the most transformed ecosystems. As illustrated in Table 7, between 1990 and 2015 forest gain 7893 ha while lose 31,820 ha, corresponding to a net loss of −23,927 ha; woodland gains 28,623 ha and lose 79,766 ha corresponding to a net loss of −51,143 ha followed by wetland which experienced the gain of 2419 ha and lose 15,956 ha equivalents to a net loss of −13,537 ha. Cultivated land, grassland, bushland and built up area was found to have consistently net gain of 36,690 ha, 30,014 ha, 38,370 ha and 4356 ha respectively.

Land covers transition matrix present transformation of each land use category. The numbers in brackets indicates the cover area which remained unchanged and others number indicate the flow of cover that changed from one cover to other cover categories. The results have reviled that all land cover categories changed with varying magnitudes. Major changes that have been noted between 1990 and 2015 include: 16,155 ha of forest was transformed to cultivated woodland and 6035 ha to cultivated land. 23,937 ha of woodland was changed to bushland; 19,510 ha to grassland and 11,588 ha to cultivated land while 4235 ha of wetland was transformed to bushland, 3016 ha to grassland and 2635 ha to cultivated land. Results implying there is encroachment of forest, woodland and wetland by human activities including agriculture.

FR: Forest, PL: Plantation, RF: Riverine forest, WTR: Water, WET: Wetland, WD: Woodland, GR: Grassland, WR: Wooded rock, CW: Cultivated woodland, BS: Bushland, CLT: Cultivated land, BLT: Built up. Numbers in brackets indicate cover areas that remained unchanged between the two periods of 1990 and 2005.

FR: Forest, PL: Plantation, RF: Riverine forest, WTR: Water, WET: Wetland, WD: Woodland, GR: Grassland, WR: Wooded rock, CW: Cultivated woodland, BS: Bushland, CLT: Cultivated land, BLT: Built up. Numbers in brackets indicate cover areas that remained unchanged between the two periods of 2005 and 2015.

EF: Forest, PL: Plantation, RF: Riverine forest, WTR: Water, WET: Wetland, WD: Woodland, GR: Grassland, WR: Wooded rock, CW: Cultivated woodland, BS: Bushland, CLT: Cultivated land, BLT: Built up. Numbers in brackets indicate cover areas that remained unchanged between the two periods of 1990 and 2015.

Summary on changed and unchanged cover areas between 1990 and 2015 are presented in Table 11. The percentage changed indicates the percentage area of a particular cover which changed to other covers while the percentage unchanged represents the percentage area of the original area of a particular cover which remained unchanged for a given period (Kashaigili & Majaliwa, 2010). Results shows, the forest cover changed to other forms by 80%, while woodland changed by 72%, cultivated woodland by 71%, wetland and water bodies changed by 83% and 66% respectively, bushland changed by 69%, grassland by 57%, built-up land and cultivated land changed by 87% and 61% respectively. However, some cover areas remained unchanged between 1990 and 2015.