Enhancing modelling of land-use and land-cover change and its impact on surface water quality in the Bangweulu sub-catchment, Zambia.
Date
2025
Authors
Lesa, Chundu Misheck
Journal Title
Journal ISSN
Volume Title
Publisher
The University of Zambia
Abstract
Wetlands are among the most productive natural ecosystems in the world, providing essential ecosystem services such as water that benefits human health, supporting aquatic ecosystems, and facilitating economic activities. However, a 64% to 71% of wetlands have been lost globally since 1900, primarily due to changes in land use and land cover (LULC). The Bangweulu Wetland System (BWS) in Zambia faces similar challenges, combined with a lack of comprehensive literature regarding LULC changes and their impacts on surface water quality. Traditional methods of LULC classification can be complex and diverse, but non-parametric approaches, such as Machine Learning (ML), have demonstrated greater accuracy. Different ML models possess distinct strengths and weaknesses, and combining multiple models has the potential to enhance the accuracy of LULC classification. Monitoring various water quality parameters in open water bodies presents significant challenges, resulting in gaps in available data. While the Water Quality Index (WQI) integrates various water quality parameters, its spatial application has not yet been thoroughly evaluated. Although, advancements in Remote Sensing and GIS technology provide wider spatial data coverage, there is a lack of comprehensive
literature regarding the impact of LULC on WQI. Additionally, most existing assessment methods have limitations, as they overlook the spatial distribution and proximity of LULC pollution sources to the water body areas of interest. Therefore, the specific objectives of this study were; (i) to ensemble a superiorly hybrid machine learning model for enhanced accuracy of modelling LULC changes, (ii) to rapidly evaluate the water quality variability, and to investigate the influence of LULC on water quality. To classify LULC, six ML models were employed: Support Vector Machine (SVM), Naive Bayes (NB), Decision Tree (DT), Artificial Neural Network (ANN), Random Forest (RF), and K-Nearest Neighbour (KNN). The study
analysed Landsat 8 images from 2020 and Landsat 5 images from 1990, 2000, and 2010 using QGIS. To establish an integrated remote sensing approach for monitoring of water quality, the study utilised Sentinel 2 data alongside on-site and laboratory water quality measurements. This integration facilitated the transformation of water quality parameter maps into WQI maps, which were then combined to present an overall assessment of water quality in the BWS lakes. Additionally, a parametric Weighted Inverse Distance Function (WIDF) was applied to determine
the contamination effective contribution area (Aec) for each LULC. This analysis utilised the classified Landsat 2020 image, field water quality data and a 30m Digital Elevation Model (DEM). A multiple regression analysis was employed to explore the relationship between Aec and specific water quality parameters, as well as the WQI. Results revealed that four models— SVM, NB, DT, and KNN—outperformed the other models. Consequently, a hybrid model, referred to as the Quad model, was developed by integrating the outputs of these four models. This Quad model showcased superior performance compared to individual models, achieving Kappa Index scores of 0.87, 0.72, 0.84, and 0.87 for the years 1990, 2000, 2010, and 2020,
respectively. The analysis of LULC changes from 1990 to 2020 indicated a yearly decline of 1.17% in forest coverage, -1.01% in grassland, and -0.12% in water bodies. In contrast, built-up areas and cropland increased at rates of 1.70% and 2.70%, respectively. Water quality assessments showed that the mean WQI from on-site and laboratory data was 34.948, while remotely sensed data yielded a mean WQI of 40.633. Both on-site/lab and remote sensing methods indicated that the concentration of water quality parameters in the Bangweulu Wetland lakes is lower (better) than the local and international recommended limits, with a calculated WQI falling within the 'Good' category. This suggests that the water is generally fresh, clean, and suitable for various uses, including ecological preservation, agriculture, aquaculture, recreation, industrial applications, and human consumption. The study also highlighted significant correlations between LULC and water quality parameters. Turbidity, TDS, iron (Fe²⁺), and EC exhibited strong correlations with specific LULC types, particularly built-up and forested areas. Conversely, parameters such as potassium, sodium, chloride, and calcium showed weak correlations with LULC. The WQI itself demonstrated a reasonable correlation with LULC (R²
= 0.649). The findings underscore the consistent growth of cropland and built-up areas from 1990 to 2020, alongside a reduction in forest cover and grassland. Although the water body experienced a gradual decrease over this period, the decline was minimal. Long-term monitoring will be essential for evaluating the success of interventions, guiding conservation efforts, and determining whether the reduction in water bodies is a sustained trend or a short-term phenomenon. This information is crucial for developing sustainable LULC policies, identifying hotspots of potential water quality degradation, and targeting areas for restoration efforts, with significant implications for future management practices.
Description
Thesis of Doctor of Philosophy in Integrated Water Resources Management