University of Twente Student Theses


Mapping crops in smallholder farm systems from high-spatial resolution and multi-temporal satellite images

Mukunga, Dickson Nzumbi (2020) Mapping crops in smallholder farm systems from high-spatial resolution and multi-temporal satellite images.

Link to full-text:
(only accessible for UT students and staff)
Abstract:Crop mapping in smallholder farm systems has been hindered by the unavailability of high spatial resolution images. Sentinel-2 images have provided a sigh of relieve for mapping crops in smallholder farms systems due to improved spatial resolution. The emergency of PS images has presented new and attractive data product to maps crops in smallholder farm systems. Thus, the study makes use of PS images to map crop types in smallholder farms systems using TWDTW and tcDTW. The two classifiers were trained using the same data for comparison of the classification results. The comparison between the resulting overall classification accuracies from the two classifiers was tested using McNemar’s test. The usability of the PS images was evaluated for spatial, spectral and temporal resolutions. The object-based classification results showed a slightly higher overall accuracy (78.08%) than pixel-based classification with an overall accuracy of 75.78%. McNemar’s Chi-square test showed that the results of the two classifications were statistically significant different. Spatial suitability of PS images made use of the segmentation goodness. The spectral suitability made use NDVI generated from red edge bands of S2 images. The temporal resolution evaluation of PS images made use of the temporal nature of PS images. It was concluded that the object-based classification produced better classification overall accuracy and a more homogeneous crop type map. Additionally, from the vegetation indices generated from the red edge bands, NDVIREA1, NDVIREA2, and NDVIREA4 produced better overall accuracy than classical NDVI hence can be used to improve classification accuracy. Contrary, the other vegetation indices generated from red edge bands (NDVIREA3, NDVIren1, NDVIren2, NDVIren3, NDVIre1, NDVIre2 and NDVIre3 resulted to overall accuracy equal to or lower than the one obtained using classical NDVI hence did not improve the classification accuracy.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page