**5. Discussion**

### *5.1. Quality of Results from Composite Image Classifications*

During this study, data from Landsat 5, 7, and 8 archives were used to form different image composites, the supervised classifications of which (under the GEE platform) led to the production of land cover maps of Togo. Despite difficulties that were encountered in finding the best quality images, the results that were obtained indicate relatively high overall accuracies of 91% to 98% for composites with the original bands and 86% to 96% for those including the vegetation indices. However, the classification results including vegetation indices tended to overestimate the built-up and bare land (buildings + soil) class and the water body class. We believe that this is likely due to the simultaneous presence of NDBI, which captures residential areas and bare soil, the BSI, which is a bare soil-specific index, and NDWI, which would have difficulty distinguishing water bodies from shadows. These results are consistent with those of [78] and [24], who found that the NDBI and modified NDWI yielded image classification results with very low accuracies, despite being two popular indices in the literature.

The results have shown that OA and *K*s for the original composite band classifications are significantly different from those with vegetation indices, but the latter did not improve the image classification results as one would have expected. Nevertheless, the spontaneous decrease in overall accuracy and *K*s for the 2005 composite classification (Tables 4 and 5) could be primarily related to deficiencies in the Landsat 7 data that are observed as fine stripes on the 2005 map (Figure 9). It should be noted that this sensor suffered hardware failure in its Scan Line Corrector (SLC) in 2003, resulting in the loss of about 22–25% of the data in each scene [79]. Additional research could be done on the impact of these indices on the quality of image classification results and also test new indices such as the Emissivity Contrast Index (ECI), which have overcome the NDVI limitation concerning its capability to distinguish bare soil from senescent vegetation [80]. Another thing that could be tested in future research using RF in order to improve image classification accuracy is to tune the hyper-parameters of this model to improve its performance [65], instead of using the default number of trees.
