Introduction

The sensitivity of watersheds or basins to human-induced alterations on land and the associated hydrological implications remains a subject of scientific interest, especially in African basins where it is often misunderstood due to limited research. While extreme hydrological events are invariably attributed to climate change, that may only be one among two important factors, with land use or land cover change (LULCC) being the other. Several theories regarding LULC change impact on hydrology, vis-à-vis the role of forest vegetation, have over the years been advanced. Among them; (1) the “sponge effect” theory which suggests that forests enhance dry-season flow by improving the soil macro-porosity and hydraulic conductivity and (2) the “infiltration evapotranspiration trade-off” which postulates that both, deforestation or reforestation can cause either a net gain or loss on the groundwater recharge and baseflow depending on whether a particular change in cover influences more on infiltration gains or evapotranspiration losses (Krishnaswamy et al. 2013). It is not uncommon therefore that studies relating forest removal to hydrological processes have over the years presented evolving findings, and at times conflicting.

For instance, experimental studies by Hibbert (1965), Bosch and Hewlett (1982), Zhang et al. (1999), and Andréassian (2004) among others, indicated that a reduction in forest vegetation causes an increase in mean annual streamflow. Hibbert (1965) even suggested that in a significantly wet watershed, the increase in water yield is proportional to the reduction in forest cover. However, Ring and Fisher (1985) analysed rainfall–runoff response in a catchment in New South Wales—Australia where 45% of native forest was replaced by dryland cropping and improved pastures; their study reported a reduction in runoff, contrary to the widely reported trend towards an increase in runoff following clearing of forest vegetation. Wilk et al. (2001) in another study failed to detect any hydrologic changes despite a reduction in forest cover from 80 to 27% in a tropical monsoon catchment (12,199 km2) where the average annual rainfall of 1100 mm in northeast Thailand. These uncertainties in hydrologic responses make localized studies essential; especially in the sub-Sahara Africa where most watersheds are undergoing significant transformations in LULC due to various socio-economic transitions (Olorunfemi et al. 2021).

For Malawi in particular, substantial changes in LULC, often linked to population pressure have over the past few decades been experienced (Nkhoma et al. 2021). As the population continues to rise rapidly (GoM 2018), there is inadvertent agricultural expansion and deforestation; a scenario that is worsened by the country’s overreliance on rainfed subsistence agriculture and dependency on provisioning ecosystem services for the people’s livelihoods (FAO 2011). It is estimated that between the years 2010 and 2020, cropland expansion alone in Malawi increased by at least 8% from 39.6% of the total country land area (Li et al. 2021). Such changes in LULC, apart from threatening the existence of critical ecosystem services (Almaw et al. 2020; Foley 2005) also pose a threat to the flood frequency (Brath et al. 2006; Crooks and Davies 2001), flood severity (De Roo et al. 2000), baseflow and dry-season flow (Wang et al. 2011) and the annual mean discharge (Costa et al. 2003) of streams or rivers.

Some studies have shown that the influence of vegetation cover change on hydrological regime is most noticeable in smaller watersheds compared to large ones (Siriwardena et al. 2006). “Small” referring to watershed sizes where whole area manipulation is feasible for study purposes. For example, the paired watersheds used in the study by Zhang et al. (1999), which all reported an increase in the mean annual yield following forest reduction, had a mean size of 1.73 km2 and the majority of those were less than 1 km2 in size. On the contrary, the studies by Wilk et al. (2001) which were conducted in four large catchments showed that only two depicted similar links that consistently exist in the small ones. The failure to establish links in the case of large watersheds is attributed to the non-uniform land use change (Allan 2004), mosaics in the vegetation at different stages of regeneration (Best et al. 2003) and the higher spatio-temporal variations in rainfall (Wilk et al. 2001) that often exist over an extensive area.

While it is feasible to conduct “paired” experimental studies on small catchments, that may not be feasible on larger ones due to the practicality of manipulation. Alternatively though, time-trends (statistical) and hydrological modelling often prove to be better approaches for assessing the impacts of vegetation changes on hydrology in large catchments (Li et al. 2009). The time-trend analysis nevertheless does not consider the physical catchment processes despite being quite easy to realize; and is therefore only limited to analyzing the hydrological effects of environment change. To fully characterize the flow regime of a catchment nevertheless, it is imperative to separate the impacts of LULC change from those of concurrent climate variability and often this presents a scientific challenge which can be solved through hydrologic modelling (Lioubimtseva et al. 2005; Tollan 2002).

For example, the SWAT model was successfully employed by Zhang et al. (2012) to separate human-induced effects from those of climate change on the streamflow of Huifa River in China. That study revealed that the decreasing streamflow was more attributable to the impact of human activities through upstream water storage and not the climatic effect. Other than SWAT, there are other hydrologic models such as the hydrologic modelling system developed by the hydrologic engineering centre (HEC-HMS) of the United States Army Corps of Engineers (USACE) that can be employed to simulate rainfall–runoff. However, HEC-HMS is a lumped model which can simulate hydrological processes of a watershed (USACE 2013). Lumped models do not consider a sub-basin as a single unit and do not have the model looping functionality. For the HEC-HMS in particular, its usage is quite limited as its model code is not made publicly available for use (Yilmaz et al. 2012).

Distributed and semi-distributed models such as SWAT on the other hand provide a more realistic framework for conceptualization of hydrologic processes (Jothityangkoon et al. 2001). Distributed models simulate land surface characteristics to a greater detail by use of model parameters as they consider each sub-basin as an individual unit (Legesse et al. 2003). Our study therefore employed the SWAT model in order to separate and quantify the implications of LULC change on the hydrological regime of four adjacent upper river catchments within the Lake Chilwa Basin in Southern Malawi. This was after an in-depth LULC change analysis was done for each of the four watersheds of the basin. Encompassed by a large wetland and rivers, the Lake Chilwa Basin is vital for its provisioning ecosystem services which benefit more than 2 million inhabitants through fishing, rice farming and potable water supply. In recent years, however, the basin has experienced recurrent hydrologic shocks that have led to major lake recessions and some rivers in the basin have shown a significant declining streamflow trend (Kambombe et al. 2018), something that threatens both water and food security and the livelihood of the people.

This study therefore aimed at providing a scientific understanding of the hydrological responses to the influences of land use/cover transitions and climate within Lake Chilwa Basin. The study builds upon a previous study that focused much on the spatio-temporal patterns of drought propagation and their influence on lake recessions (Kambombe et al. 2021).

Study site

The four studied catchments of Domasi, Mulunguzi, Likangala and Thondwe lie within the Lake Chilwa Basin of Malawi (Fig. 1). These catchments essentially drain into an important endorheic lake (Chilwa) which is a source of fish, providing food and livelihood to fishermen and traders from surrounding areas. In terms of climate, the Lake Chilwa Basin is categorized to be within the tropical wet and dry “savanna” climate. The basin experiences a unimodal rainfall pattern, with a typical season starting from November to April. The average annual rainfall ranges between 1100 and 1600 mm, but can reach up to 2500 mm in mountainous areas and can be as low as 700 mm in lowland plains (Mvula et al. 2014). Seasonal rains are mostly influenced by the converging of the north easterly monsoon and south-easterly trade winds which create a low-pressure area known as the Inter-Tropical Convergence Zone (ITCZ). However, there are sporadic winter rains drizzling mostly in highlands between the months of May and August. These rains are influenced by an influx of cool moist south-easterly winds coming from Mozambique locally known as “Chiperone” winds (Ngongondo et al. 2011). The temperatures in the basin are within mean annual ranges of 21 and 24 °C (Chavula 2000). Table 1 presents a summarized long-term average weather record from some stations used in this study, as presented by Kambombe et al. (2021).

Fig. 1
figure 1

Map showing study watersheds located within the Lake Chilwa Basin of Malawi

Full size image
Table 1 Summary information on weather stations used in the study (1970 to 2020)
Full size table

Methods and data

Quantifying changes in land use or cover

In order to examine the changes in LULC that occurred within the river catchments, land cover images from the GlobeLand30 dataset were extracted for the years 2000, 2010 and 2020. The GlobeLand30 datasets were preferred because of their relatively high resolution (30 m) compared to other land cover datasets (e.g. 250 m ESACCI, 500 m MCD12Q1). These datasets are also known to accurately represent land cover in Africa (Jacobson et al. 2015; Samasse et al. 2018; Wei et al. 2020) including Malawi. Our study adopted major land cover classes and the same hierarchical classification scheme (Chen et al. 2014; Sulla-menashe et al. 2011) as GlobeLand30 2010 (Chen et al. 2017) to ensure consistency for the 2000, 2010 and 2020 land cover maps.

The six classes included barren lands and artificial surfaces (bare land), cropland, forestland, grasslands, shrublands and water bodies; selected based on predominance in the study area and were identified in sequence. The global surface water dataset was used (Pekel et al. 2016) to identify the permanent water bodies in Malawi in 2020. Thereafter, the newly developed bare land and settlements were classified for the years 2000, 2010 to 2020 using the Enhanced Vegetation Index (EVI) (Liu and Huete 1995) threshold classification. A threshold of maximum seasonal EVI of 0.3, which classify 97% of the non-vegetated area was defined by sampling vegetated pixels (including shrubland, cropland and grassland) and non-vegetated pixels (including bare land and settlements) based on Google Earth high-resolution images. We further used the Global Forest Change dataset (Hansen et al. 2013) that has been updated to recent years to define forest area in 2020. A total of 33 training features including six spectral bands; four spectral indices; elevation, slope and aspects (Farr et al. 2007) were used to train the random forest classification algorithm.

Spectral indices that included the Normalized Difference Vegetation Index (NDVI) (Defries & Townshend 1994), Enhanced Vegetation Index (EVI) (Liu and Huete 1995), Normalized Difference Water Index (NDWI) (Gao 1996) and Normalized Difference Built-up Index (NDBI) were also calculated. Training samples of each class were selected manually by referencing recent Google Earth satellite images of Malawi and Sentinel-2 satellite RGB images. Ground-truthing exercise by physically verifying features that were not very clear on the image was also done. We used 70% of samples to train the hierarchical classification scheme, while the remaining 30% for assessing the accuracy of the final 2020 land cover classification. The final accuracy reached up to 82% with cropland accuracy of 83.5%.

LULC change transition analysis

The detection of changes that occurred between two different periods employed a post-classification comparison (PCC) change detection method. This method enables a detailed LULC change analysis of independently classified maps (Jensen 2005). PCC is a common technique used to compare maps sourced from different periods. The approach provides a comprehensive and detailed “from-to” LULC change information as it does not require data normalization between the two dates (Aldwaik et al. 2012; Coppin et al. 2004; Jensen 2005; Teferi et al. 2013). The use of the PCC technique resulted in a cross-tabulation matrix (LULC change transition matrix) which was computed using overlay functions in ArcGIS. Gross gains and losses were also calculated for two periods, i.e. 2000–2010 and 2010–2020. A summarized framework for the LULC classification processes is presented in Fig. 2.

Fig. 2
figure 2

Framework for the LULC image classification and post-classification comparison

Full size image

SWAT calibration and validation

For the hydrological modelling, the required SWAT input datasets such as land use/cover maps, soil map, DEM and climatic data were prepared into the acceptable formats and a new SWAT project was built for each of the studied catchment. Each catchment underwent the process of catchment delineation, creation of sub-watershed and the hydrological response units (HRUs). To ensure a satisfactory performance of the model, calibration and validation were both performed. The predictive power of model performance was later assessed using indicators such as Nash–Sutcliffe (NS), coefficient of determination (R2), Per cent bias (PBIAS) and the standard deviation ratio (RSR) ( Li et al. 2009; Moriasi et al. 2007). The equations are as shown:

$$ {\text{NS}} = 1 - {{\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - Q_{{{\text{sim}}}} )^{2} } } \mathord{\left/ {\vphantom {{\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - Q_{{{\text{sim}}}} )^{2} } } {\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - \overline{Q}_{{{\text{obs}}}} )^{2} } }}} \right. \kern-0pt} {\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - \overline{Q}_{{{\text{obs}}}} )^{2} } }}\, $$
(1)

where Qobs and Qsim represent the observed and simulated discharge data, respectively. The NS statistic can fall within the range of − ∞ to 1. When the NS value is between 0 and 1.0, the prediction is considered to be acceptable. The value of 1.0 denotes a perfect prediction while values less than 0 are considered unacceptable (Krause and Boyle 2005).

$${R^2} = {{{{\left[ {\sum\limits_i {\left( {{Q_{{\text{obs}}}}_i - {{\overline Q }_{{\text{obs}}}}} \right)\left( {{Q_{{\text{sim}}}}_i - {{\overline Q }_{{\text{sim}}}}} \right)} } \right]}^2}} \mathord{\left/ {\vphantom {{{{\left[ {\sum\limits_i {\left( {{Q_{{\text{obs}}}}_i - {{\overline Q }_{{\text{obs}}}}} \right)\left( {{Q_{{\text{sim}}}}_i - {{\overline Q }_{{\text{sim}}}}} \right)} } \right]}^2}} {\sum\limits_i {{{\left( {{Q_{{\text{obs}}i}} - {{\overline Q }_{{\text{obs}}}}} \right)}^2}\,\sum\limits_i {{{\left( {{Q_{{\text{sim}}i}} - {{\overline Q }_{{\text{sim}}}}} \right)}^2}} } }}} \right. \kern-\nulldelimiterspace} {\sum\limits_i {{{\left( {{Q_{{\text{obs}}i}} - {{\overline Q }_{{\text{obs}}}}} \right)}^2}\,\sum\limits_i {{{\left( {{Q_{{\text{sim}}i}} - {{\overline Q }_{{\text{sim}}}}} \right)}^2}} } }}$$
(2)

The R.2 values range from 0 to 1, with 0 indicating no correlation while 1 denotes perfect correlation (Abbaspour 2014)

$$ {\text{PBIAS}} = {{\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - Q_{{{\text{sim}}}} )} } \mathord{\left/ {\vphantom {{\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - Q_{{{\text{sim}}}} )} } {\sum\limits_{i = 1}^{n} {Q_{{{\text{obs}}}} } }}} \right. \kern-0pt} {\sum\limits_{i = 1}^{n} {Q_{{{\text{obs}}}} } }} $$
(3)

For the PBIAS, the value of 0 represents best simulation performance. When PBIAS values are close to 0, the simulation is considered satisfactory. When PBIAS is negative, it implies there is an overestimation bias while positive values show underestimation bias.

$$ {\text{RSR}} = {{{\text{RMSE}}} \mathord{\left/ {\vphantom {{{\text{RMSE}}} {{\text{STDEV}}_{{{\text{obs}}}} }}} \right. \kern-0pt} {{\text{STDEV}}_{{{\text{obs}}}} }} = {{\sqrt {\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - Q_{{{\text{sin}}}} )^{2} } } } \mathord{\left/ {\vphantom {{\sqrt {\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - Q_{{{\text{sin}}}} )^{2} } } } {\sqrt {\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - \overline{Q}_{sim} )^{2}_{{}} } } }}} \right. \kern-0pt} {\sqrt {\sum\limits_{i = 1}^{n} {(Q_{{{\text{obs}}}} - \overline{Q}_{sim} )^{2}_{{}} } } }} $$
(4)

When RSR is equal to 0, it shows that there is no residual variability and therefore the simulation is perfect (Li et al. 2009). Smaller values of RSR imply better performance for the model.

Analysis of LULC change impact on hydrology

LULC change scenario runs

In order to assess the impact of land use change on the hydrological regime, three scenarios; S1, S2 and S3, as described in Li et al. (2009) were independently and subsequently run using land use maps for the year 2000, 2010 and 2020, while holding constant all such factors as climate data and model parameters. The simulated results were then used to compare hydrological effects of land use change on streamflow behaviour in the four study watersheds of Domasi, Likangala, Mulunguzi and Thondwe. Table 2 shows the setup combinations of scenarios.

Table 2 Scenario setup for the climate variability and land use change simulations
Full size table

Probability of exceedance computations

The probability of exceedance was then computed using the daily output streamflow data obtained from the LULC scenarios, for each of the four watersheds. The computations considered the long-term record of streamflow as percentiles at an interval of 0.5 and the FDC 2.1 Hydro Office programme was used. As in Siriwardena et al. (2006), the flow duration curve plots were made as log plots on the same axis for the 3 land uses for comparison of runoff behaviour. This enabled the understanding of runoff characteristics exhibited as a response to land use change for each watershed.

Assessing LULC change impact on evapotranspiration

Evapotranspiration (ET) and soil water (SW) as a function of climate variability and land use change were also estimated using SWAT. The simulated average annual SW and annual ET from the three scenarios (S1, S2, S3) were also compared. The Hargreaves method of estimating the potential evapotranspiration (PET) and actual evapotranspiration (ET) was employed, selected particularly because of its less data demands compared to other techniques such as the Penman–Monteith. The Hargreaves equation (Hargreaves et al. 2003) is given by:

$$ {\text{ET}}_{0} = 0.0023{\text{Ra}}(T_{\max } - T_{\min } )^{0.5} \left[ {(T_{\max } + T_{\min } )/2 + 17.8} \right] $$
(5)

where ETo denotes reference evapotranspiration (mm day−1), Ra represents the extraterrestrial radiation (MJ m−2 day−1) while Tmax and Tmin represent the maximum and minimum air temperatures (°C), respectively. The ET in this case was estimated taking into account the influence of factors such as ground cover and canopy properties (Jovanovic and Israel, 2012). ET is given by:

$$ {\text{ET = }}k_{c} + {\text{ET}}_{{\text{o}}} $$
(6)

where kc represents the crop coefficient

Assessing LULC change impact on soil water

Soil water (SW) in SWAT was estimated using the water balance formulation as explained by Neitsch et al. (2011) and is given by:

$$ {\text{SW}}_{t} = {\text{SW}}_{0} + \sum\limits_{i = 1}^{t} {\left( {R_{{{\text{day}}}} - Q_{{{\text{surf}}}} - E_{a} - w_{{{\text{seep}}}} - Q_{gw} } \right)} $$
(7)

SWt denotes the final SW content (mm), SW0 represents the initial content of SW on day i (mm), t indicates time (days), Rday being precipitation amount on day i (mm), Qsurf represents the surface runoff on day i (mm), Ea is the evapotranspiration on day i (mm), Wseep represents water amount from the profile seeping through the vadose zone on day i (mm), and Qgw is the return water flow on day i (mm).

Assessing climate variability impact on streamflow

For the estimation of climate variability impact on river flow, climate data (rainfall and temperature) were divided into decadal blocks of 1980 to 1989, 1990 to 1999, 2000 to 2009 and 2010 to 2009 representing the 1980s, 1990s, 2000s and 2010s. Scenarios runs were then made using each decadal climate data subsequently while holding constant all other factors such as land use. Results from each scenario were then compared and plotted.

Results and discussions

Changes in land use or land cover

Figures 3 and 4a to d are the results of the general pattern of LULC change in the four studied watersheds. As shown, the general pattern was that of increasing cropland area and declining forest vegetation. This is a general reflection of the changes in land cover at country level, as supported by findings by Li et al. (2021) who reported a rapid overall cropland expansion of 8.5% between 2010 and 2019 for Malawi.

Fig. 3
figure 3

The proportion of major LULC for each watershed and year analysed

Full size image
Fig. 4
figure 4

Land use or cover maps for each watershed in the years 2000, 2010 and 2020

Full size image

The results reveal that the extent of LULC change varied significantly from watershed to watershed, and such variations presented a study opportunity for teasing out hydrologic responses under varying conditions. For instance, the Domasi Watershed, with a total land area of 67.8 km2 had a significant decline in forest area by 39.3% from 36.1 km2 to 21.6 km2 between 2000 and 2010, representing a 21% net loss of the total watershed area. However, between 2010 and 2020 there was only a 2% net loss in forest area within Domasi.

Mulunguzi Watershed on the other hand, with the smallest area of 20 km2 compared to all study watersheds had the biggest proportion of area under forest at 93% in 2000 which reduced to 81% by 2010 and 61% by 2020, representing a decline of 12 and 19%, respectively. The other two watersheds of Likangala (76.1 km2) and Thondwe (276.4 km2) though had a marginal change in LULC. These two watersheds were dominated by cropland, particularly the Thondwe Watershed with 87% under cultivation in 2000 which increased to 94 and 95% by the year 2010 and 2020, representing an increase in 7 and 1%, respectively. At the same time, forest area declined by 3.9 and 0.1%. Figure 5a to d shows the net gain or loss between periods for each of the major LULC classes that include cropland (CL), forestland (FL), grassland (GL), shrubland (SL), waterbodies (W) and bare land (BL) for the four watersheds.

Fig. 5
figure 5

Net gain or loss in LULC for each watershed for the study period

Full size image

SWAT model performance

Model performance was satisfactory as shown by the coefficient of determination (R2) values of 0.90 and 0.89 obtained after calibration (1983–1987) and validation (1988–1992) periods, respectively (Fig. 6a and b), using the Thondwe Watershed. These values indicate a strong relationship that existed between the observed and simulated flows, both in the calibration and validation periods. Furthermore, the seasonal plot of simulated and observed discharge for both the calibration and validation periods in Fig. 7 shows the model performed satisfactorily.

Fig. 6
figure 6

Goodness of fit for the calibration and validation of the model

Full size image
Fig. 7
figure 7

Observed and simulated mean monthly discharge for the calibration and validation periods

Full size image

Model performance indicators, which include the NS, PBIAS and RSR statistics are tabulated in Table 3. As shown, NS values of 0.88 and 0.86, PBIAS values of − 11.6 and − 19.8 and RSR values of 0.34 and 0.37 were achieved after calibration and validation, respectively. These statistics indicate good model performance for both periods of calibration and validation. Moriasi et al. (2007) stated that a model simulation is considered satisfactory if; NS > 0.5, PBIAS <  ± 25% and RSR ≤ 0.7. Most studies, however, only consider two statistics, the R2 and NS as adequate for the assessment of model performance (Arnold et al. 1998).

Table 3 A summary of model performance results after calibration and validation
Full size table

Impact of LULC change on hydrology

Figure 8a to d presents plots of the flow duration curves (FDCs) produced from the scenario outputs. Since the climatic factor (i.e. rainfall, temperature, solar radiation, etc.) was held constant during the consideration of LULC change impact in all the watersheds, any resultant change in the hydrological regime was attributable to the land surface alterations.

Fig. 8
figure 8

Flow duration curves (FDC) under different land use conditions in the study watersheds

Full size image

From the FDC plots, the Domasi Watershed (67.8 km2) clearly presents a notable increase in runoff response to the changing LULC from the year 2000 to 2010. This increased runoff tendency is linked to the alteration of surface flow regimes caused by a significant reduction in forest cover by a net loss of 21%, which was replaced mostly by maize cultivation (15%). Field crops such as maize, having undergone different stages of crop development within a season contribute less to rainfall interception compared to natural vegetation or forests; hence more overland flow or runoff is generated in field crops particularly with poor farming practices.

Natural “tropical” forests enhance groundwater recharge through an intertwine of litter, soils and roots which provides a “sponge” effect that enhances water seepage during the rainy season releasing it gradually during the dry period (Bruijnzeel 2004). Deforestation on the other hand tends to disrupt the “sponge” effect through compaction due to raindrop impact and grazing. Again, there is rapid oxidation of soil organic matter where the soil is exposed (Grip et al. 2005). Because the net loss in forest cover was only 2% within the Domasi Watershed from 2010 to 2020, there was no notable change in runoff response as presented in the FDCs. Although grassland significantly decreased by a net loss of 16% to possibly influence a change in runoff pattern, the net gain in shrubland area by 13% during the same period effectively countered any possibility for a change in runoff response. Shrublands have a relatively high rainfall interception potential compared to most cultivated crops (Costa et al. 2003; Kambombe et al. 2018); mostly because at the initial stage of crop development, the land is almost bare and therefore exposed to more overland flow.

For the watersheds of Likangala (76.1 km2) and Mulunguzi (20 km2), a similar increasing runoff response pattern as that displayed in the Domasi Watershed was also portrayed; except that the extent of change was not that much pronounced. For the Likangala Watershed, the reasons for the dismal change in runoff behaviour as shown in Fig. 8b are attributed to a fairly small margin of change in LULC, with forest cover having a net loss of 7% between 2000 and 2010, and a 2% loss between 2010 and 2020. Unlike Likangala, the Mulunguzi Watershed experienced a higher margin of land cover change despite registering only a dismal increase in runoff response (Fig. 8c). For example, there was a net loss 12% forest cover from 2000 to 2010, while grassland and shrubland had a net gain of 8% and 3%, respectively. The gains in shrubland area effectively minimized the runoff generation. This was mostly the case in 2020 whereby despite the watershed registering a net loss of 19% in forest cover compared to 2010 land use map, there was a concurrent 29% net gain in shrubland area which adequately contributed to rainfall interception and reduced runoff.

For the Thondwe Watershed which had the biggest watershed area (276.4km2) of the four, the runoff generation response to the changing LULC was almost negligible as shown in Fig. 8d. This was largely because Thondwe was already predominantly under agriculture land for the study period, and there was quite a small margin of LULC change, with the biggest margin of change being a 7% increase in cultivation between 2000 and 2010. Nevertheless, irrespective of the minimal change in LULC for the individual classes, the possible combined effect of increased cropland area (+ 7%), declined forests vegetation (− 4%) and declining grassland area (− 3%) between 2000 and 2010 could suggest a possible noticeable increase in runoff response. Our failure to detect a significant change could be attributed to the large size of the watershed. Other studies observed that the bigger the watershed size, the more difficult it is to detect any hydrologic response associated with LULC change (Best et al. 2003; Wilk et al. 2001). In any case, the effectiveness of forest canopy cover in controlling runoff generation can, however, only be achieved at certain threshold level of cover (Zhao et al. 2014), which clearly was not satisfied in the case of Thondwe Watershed.

Table 3 presents more hydrological response results based on the study scenarios. Parameters of interest were the runoff depth (RO), soil water (SW) and evapotranspiration (ET) presented as depth in mm; and the per cent change in response to the changing LULC.

The simulated runoff (RO) response depth is generally a reflection of the FDCs in Figs. 8a to 8d. Domasi watershed for example presents the biggest margin of runoff response of 11.4% from 2000 to 2010. The Likangala and Mulunguzi Watersheds also demonstrated a notable margin of runoff response to the changing LULC from 2000 to 2010 of 7% and 8%, respectively. However, Thondwe presented the least response at 3% over the same period. From 2010 to 2020, all the watersheds present the smallest margins of change just as in the FDCs.

In terms of the soil water (SW) component, the general depiction in all the four watersheds is that of decreasing SW in response to the changing land use or cover, although the degree varied. This could be linked to the increasing runoff tendency mostly caused by deforestation and increased cropland expansion that dominated most of these watersheds. As runoff increases, there is more overland flow and less infiltration leading to decreasing SW and groundwater recharge. This in turn often negatively affects the dry-seasonal flows of rivers and streams. Based on the study by Kambombe et al. (2018) which analysed baseflow trends of major rivers within the Lake Chilwa Basin, it was revealed that there is a declining trend in the dry-season flows especially in the most degraded watersheds of Domasi, Likangala and Thondwe. Declining baseflows imply there is an increased risk of these pereneal rivers and streams becoming seasonal or ephemeral, which poses a threat to water and food security in the basin.

Regarding the evapotranspiration (ET) component, water loss through ET was quite high in such watersheds where vegetation and temperatures were also significantly high, as in the Domasi where almost half of the annual precipitation was shown to be lost as such. Although ET was quite high in those watersheds, the margin of change in the ET component, as a response to the changing LULC between periods, was quite small compared to that exhibited by the runoff component; with the Domasi Watershed presenting the biggest margin of ET per cent change at -1.6% as a reflection of the change in LULC from the year 2000 to 2010. Since the ET response is very small compared to that displayed by the RO response, it implies therefore that the positive contribution forests make in controlling RO within these watersheds surpasses the negative contribution they make in terms of water loss through ET. Our findings therefore affirm that the “sponge effect” role of forest vegetation is than the “infiltration-evapotranspiration tradeoff” within the Lake Chilwa Basin.

Climatic influence on streamflow

Figure 9 presents the inter-decadal relationship between rainfall and discharge; with a focus on the Thondwe Watershed which was used in the calibration and validation of the model. As shown, there is a direct linkage between the decadal average annual rainfall and river discharge. Rainfall is a major component of climate that influences streamflow generation in Malawi, such that inter-annual rainfall variability tends to influence the inter-decadal rainfall. Malawi’s wet rainfall seasons are largely influenced by the Indian Ocean Sea Surface Temperatures (Kumbuyo et al. 2015) (Table 4).

Fig. 9
figure 9

Inter-decadal comparison of rainfall and discharge

Full size image
Table 4 Simulated average annual runoff depth, soil water and evapotranspiration under different land use but similar climate
Full size table

As shown, there is a decrease–increase–decrease trend in the decadal rainfall and streamflow from the 1980s to 2010s. Climatological changes from 1980 to 1990s resulted to a notable reduction in rainfall which in turn reduced the decadal streamflow by 18%. However, from 1990 to 2000s, climatological changes accounted for a 10% increase in streamflow, while from 2000 to 2010s a 7% decrease was observed. The notable decrease in streamflow from the 1980s to 1990s is consistent with the general climate pattern, as also reported by other researchers such as Nkhoma et al. (2021) who observed a 49.8% reduction in decadal rainfall from the 1980s to 90 s over the Wamkurumadzi River Watershed in Southern Malawi. Similarly, Coulibaly et al. (2015) and McSweeney et al. (2010) noted that climate regimes in Malawi portrayed a significant negative shift in the early 1990s.

From the plot, the decadal rainfall from the 1990s to 2000s showed a positive trend although the general long-term rainfall trend in the Lake Chilwa Basin is shown to be towards a decline. Recent findings from Kambombe et al. (2021) further show that the long-term rainfall trend in the Lake Chilwa Basin is towards a decrease, although at a statistically insignificant trend based on the modified Mann–Kendall test (Kendall 1975; Mann 1945) at α = 0.05. At country level, several studies have similarly shown that rainfall is insignificantly decreasing while temperatures are significantly increasing in Malawi (Ngongondo et al. 2011; 2015). With rainfall tending to decrease, the implication is similarly towards decreasing river flows in the basin and in Malawi broadly.

Topographic influence

Figure 10 shows slope maps of each of the studied watershed. As presented, the Domasi Watershed is shown to have the biggest proportion of steep areas with over 40% of the area being within a steep slope of > 30%. Slope is a factor in runoff generation with steeper slopes generally increasing the possibility for overland flow and reduced infiltration (Grosh and Jarrett 1994). This explains why slope is considered as a principal component in the creation of hydrological response units (HRUs) during surface hydrological modelling. The steep slopes within the Domasi Watershed may have influenced a notable runoff response. In this study nonetheless, the only changing variable in each study watershed was LULC and slope was held constant; such that any resulting change in the hydrological response in any particular watershed was attributed to the LULC change.

Fig. 10
figure 10

Map showing categorized range of slope as a percentage for each

Full size image

Conclusions

High deforestation and increased cropland expansion largely characterize watersheds within Lake Chilwa Basin, particularly between the years 2000 and 2010. However, from 2010 to 2020, there was a notable increase in shrubland area in most of the previously cleared landscapes. The increase in shrublands presents an opportunity for faster forest restoration efforts, assuming regenerated plants remained undisturbed for some years. Our study further shows that SWAT model is quite effective in modelling impacts of environmental change on surface hydrology; with model indicators such as Nash–Sutcliffe model efficiency (NS), Per cent bias (PBIAS) and root mean square error-observations standard deviation ratio (RSR) giving values of 0.88 and 0.86, − 11.6 and − 19.8 and 0.34 and 0.37, respectively, obtained after model calibration and validation. It is evident from this study that the runoff behaviour in the degraded watersheds was quite high which consequently leads to declining baseflows in such watersheds. The increasing runoff tendency nonetheless was more pronounced in smaller catchments as opposed to larger ones. Also, the study reveals that the evapotranspiration component in the basin is quite high, and that forests vegetation tended to escalate water loss through evapotranspiration. Nonetheless, the role of forest vegetation in reducing runoff and enhancing groundwater recharge is shown to surpass the ET loss, which suggest dominance of the “sponge effect” where forest cover existed. However, since most watersheds in the basin remain significantly deforested, a threat of increased runoff leading to flooding and declining dry-seasonal flows is certain.