A remote sensing based delineation of the areal extent of smallholder sugarcane fields of South Africa.

Master of Science in Environmental Science. University of KwaZulu-Natal, Pietermaritzburg 2016.

The areal extent delineation of smallholder sugarcane fields in fragmented landscapes is a challenge due to their complex spatial configuration (i.e. patchy field sizes) and timeless planting and harvesting dates. Nevertheless, delineating and estimating areas of such farming systems is essential in crop yield estimation as well as food supply inventorying to enhance food security planning for the country. Moreover, estimating the areal extent of fragmented smallholder fields can provide insights into their natural resource uses as well as their contribution to carbon pool. However, the challenge is the lack of robust, applicable methods and platforms that could be used to accurately map these farming systems in a quick, efficient and cost-effective manner. Based on that premise, this study sought to evaluate the utility of remotely sensed data coupled with advanced machine-learning classification algorithms for estimating the areal extent of smallholder sugarcane fields. The scope of this study was limited to (1) evaluating the performance of support vector machine (SVM) at pixel-based image analysis (PBIA) and object-based image analysis (OBIA) platforms in delineating areas of fragmented smallholder sugarcane fields using Landsat 8 Operational Land Imager (OLI) imagery (2) Comparing support vector machine and random forest (RF) in delineating the areal extent of smallholder sugarcane fields based on Landsat 8 OLI imagery. The performance of the two algorithms was determined based on accuracies derived using confusion matrices. Based on objective 1, the findings show no statistical significant difference (p ≥ 0.05) between PBIA and OBIA when using support vector machine (SVM). Furthermore, when comparing SVM with RF an increase of 6% was observed in overall accuracy. Nevertheless, results from the McNemar’s showed that the 6% difference was not significant. From the findings on this study, it was concluded that (1) Support vector machine can reduce the accuracy gap between PBIA and OBIA in delineating areas of smallholder sugarcane fields based on Landsat 8 OLI imagery, (2) Despite observing no statistical significance difference in accuracy, SVM outperformed RF by a margin of 7%. Meanwhile, both RF and SVM have great potential in delineating areas of the fragmented smallholder sugarcane fields.