River data on discharge and characteristics is essential for water management and water supply, as well as for flood prediction and flood control (Pan, Wang, and Xi 2016). In practice, many watersheds are ungauged due to high costs, inaccessibility and even due to political instability (Pan, Wang, and Xi 2016). For this reason, measuring remotely without the need of being physically present, for instance by remote sensing satellites, can be interesting for many applications. The large amount of satellite data can result in the ability to extend short observation series into larger series with satellite missions.
Discharge is one of the conditions in a river, which is relevant to have data on during regular periods but in particular during or after extreme events. This thesis focussed on an approach, by using remote sensing, to obtain data that can be used for further research to determine discharge. River-width is one of the current variables researched to be used as a substitute for river stage data. River stage is currently used to obtain estimations for river discharge via earlier obtained river stage-discharge relations, which can be transformed into river width-discharge relations.
The objective of this thesis was to develop a method to obtain sub-pixel accurate river-width estimations by remote sensing. The objective to estimate river-widths on sub-pixel base originates from the need of river-width estimations with higher accuracy than the freely available optical satellite resolutions of 10 to 20 metres. The study contains the improvement of the current water classification methods by including analyses for discriminating band combinations, to construct site-specific indices. This was noticed to be needed, due to the conventional indices, like the NDWI, performing differently with the presence of certain land types.
By having multiple indices based on uncorrelated satellite bands transformed into probability bands, it is possible to combine indices, via Bayes theorem. Based on the site-specific indices and index combinations, the aim is to develop relations between spectral information and water fractions of pixels that could lead to a more detailed river-width estimation by including sub-pixel information.
The resulting method was able to show discriminating abilities in satellite bands and band combinations, specifically for an area of interest, other than the conventional NDWI and MNDWI. With the use of river edge information, the spectral bands could be transformed into spatial water probability bands, indicating a probability for the present pixels to be water. The probability indices and index combinations showed to reduce a large part of the occurring misclassifications. With the use of ROC curves, to assess the classification performance of the indices and combination of indices, variation in misclassification of certain land types between days were observed for certain indices.
The probability bands, which are based on the river’s edge value distribution, also seemed to be useful, especially for the pan-sharpened MNDWI and the Bayes 0-3 indices, to obtain the needed water fraction relations for sub-pixel base estimations. A comparison in river-width estimation of a conventional automated water classification method; Otsu’s thresholding method and a Supervised training map classification method were made against the use of probability indices with sub-pixel water fraction relationships. It was found that the use of sub-pixel information resulted in a significant improvement of the accuracy for river-width estimations. For the first and second fieldwork day, the average river-width deviations of the pan-sharpened MNDWI and Bayes0-3 decreased, respectively, from 16 and 9 metres, to under 5 and 7 metre deviation by including the found water fraction relationships.