Global land cover change is characterised by the expansion of agricultural and urban areas, which results in forest loss, especially in sub-Saharan African countries such as Zambia. This topic has received increasing research attention due to the close relationship between land cover and land use, food security and climate change. The Zambian landscape has high rates of land cover change associated with deforestation, forest degradation and urbanisation. With these changes, managing natural resources in Zambia requires reliable information with which to make informed decisions during land use planning. However, existing land cover information for Zambia is limited in its spatial and temporal scales. The availability of remotely sensed data with an open access policy and a long historical record, such as Landsat satellite imagery, offers opportunities for monitoring long-term land cover change over large areas.

This thesis aims to provide an understanding of the different aspects of remotely sensed land cover monitoring, including image pre-processing, land cover classification, land cover change and factors associated with land cover change in Zambia. This research was conducted at a national scale over a period of four decades (1972–2016). The current study started with a detailed literature review on the development of the methods of Landsat land cover classification, which was followed by testing machine-learning classifiers and pre-processing methods on pansharpened and non-pansharpened Landsat Operational Land Imager (OLI-8) images. Classification of nine land cover types (primary forest, secondary forest, plantation forest, wetlands, cropland, irrigated crops, grassland, waterbodies and settlements) was conducted for six time steps (1972, 1984, 1990, 2000, 2008, and 2016), which were chosen by considering past economic and political events in Zambia. Post-classification analysis was then applied in order to understand changes in land cover. Finally, the factors contributing to land cover change were assessed using a classification tree (CT) approach.

The literature review (Chapter 2) showed that Landsat land cover classification methods have developed from manual delineation to advanced computer-based classification methods. These developments have occurred due to the advancements in computer science (e.g. machine- learning and artificial intelligence) and improvements in remotely sensed data acquisition. To attain high land cover classification accuracies, Landsat images require the selection of an effective classification method and the application of pre-processing methods. The combination of object-based image analysis (OBIA) and machine-learning classifiers, such as random forests (RF), has become more common than the pixel-based approach.

The assessment of pre-processing methods (Chapter 3) on two provinces of Zambia (Copperbelt and Central), which were covered by four Landsat OLI-8 images, indicated that applying both atmospheric and topographic correction improved classification accuracy. The results showed that non-pre-processed images reached a classification accuracy of 68% for pansharpened and 66% for standard Landsat OLI-8 images. Classification accuracy improved to 93% (pansharpened) and 86% (standard) when combined moderate-resolution atmospheric transmission (MODTRAN) and cosine topographic correction pre-processing were applied. The results showed that image corrections are more important when applied on multiple scenes, especially for time series studies. The results also identified that the RF classifier outperformed the other classifiers by attaining an overall accuracy of 96%. These results informed the choice of pre-processing and classification analyses to use for the subsequent land cover analysis.

A nationwide land cover classification analysis was then undertaken for each of the six time steps (Chapter 4). Overall accuracies ranging from 79% to 86% were attained, with more recent time steps, captured by Landsat OLI-8 imagery, having the highest accuracy. The variation in classification accuracies was mainly attributed to the differences in spatial, spectral and radiometric resolutions of the satellite images available for each time step. This chapter also showed that 62.74% of the Zambian landscape experienced change. Primary forest declined from 48% to 16% between 1972 and 2016, while secondary forest increased from 16% to 39% during the same period. The results also showed that forests have been recovering by 0.03% to 1.3% yr⁻¹ (53,000–242, 000 ha yr⁻¹); however, these rates are lower than deforestation rates (−0.54% to −3.05% yr⁻¹: 83,000–453,000 ha yr⁻¹). Annual rates of change varied by land cover, with irrigated crops having the largest increase (+3.19% yr⁻¹) and primary forest having the greatest decrease (−2.48% yr⁻¹). Area of settlements, cropland and grasslands increased, while wetlands declined. Due to increased forest fragmentation, forest connectivity declined by 22%.

The CT models for analysing the factors contributing to land cover change (Chapter 5) were produced with overall accuracies ranging from 70% to 86%. CTs are statistical approaches used to partition categorical data (response variables) into mutually exclusive subgroups using a set of explanatory variables. Here, the response variables included a binary scenario (change/no change) and changes from individual land covers. The explanatory variables were the different factors considered to be associated with the land cover changes. The major factors associated with the binary scenario (change or no change to 1972 land cover) were percentage of cultivated area, crop yield, and distance to waterbodies. Forest losses were mainly associated with crop yield, area under cultivation, population density and distance to roads and railways. An important insight from this chapter was the influence of protected areas (e.g. national forests) on forest reversion and recovery.

Due to the national extent and long temporal record of land cover change, the findings from this thesis are important for land use planning in Zambia. This research not only documents deforestation and forest degradation occurring throughout Zambia over the past four decades, but also highlights the importance of increasing the extent of protected areas in order to support forest reversion and recovery. Since forests are an important component of climate change mitigation initiatives, these results will provide baseline information for international climate change mitigation initiatives such as reducing emissions from deforestation and forest degradation (REDD+).