Review of the methods

42

43 4.1.2. Image preparation

The PRISMA images are made available with a standard georeferencing with an accuracy up to 200m.

The exact location of the in-situ samples could not be located using the GPS-grid of the image. They had to be located manually and visually using the 5m spatial resolution panchromatic images and aerial images from Google Earth that did show the correct location of sites. Based on the 30m resolution of the RGB images, it is estimated that the sites were in the imagery with an accuracy of one 30m pixel.

It is possible to receive PRISMA images with an accurate georeferencing if Ground Control Points (GCP) are available for the area (Agenzia Spaziale Italiana, 2020). Using such imagery increases the accuracy of the reflectance retrieved from pixels located over sampling sites. For images of the Danube-Sava confluence, no GCPs were available.

In the process of avoiding and removing pixels with mixed values from the analysis, some features were overlooked. In the application of the original band-ratio and NDCI algorithm, as well as the calibrated algorithm, local elevated chl-a concentrations were predicted that were concentrated to 1 to 4 pixels, As can be seen in Figure 35-38, these elevations were located where large ships, barges or bridges were present. Removing and masking out these features from the image was not achieved during pre-processing. The reflectance that these features added to the pixels’ average were such that the algorithms falsely registered these as a relatively high chl-a concentration. This phenomenon can also be caused by a large number of smaller objects. As depicted in figure 39 and 40, a large number of small, recreational vessels are anchored evenly distributed over the southern, left arm in the confluence. These vessels cause a change in reflectance values and hence, an apparent increase in chl-a concentrchl-ation chl-as well. As these phenomenchl-a occurred chl-at only chl-a few locchl-ations chl-and chl-as these locchl-ations were known, it did not affect the interpretation of the chl-a distribution patterns. The decision was made to not redo the image preparation and remove the mixed pixels in order to estimate the chl-a concentrations without these mixed pixels.

44

Figure 35 Elevated chl-a concentrations at locations of a ship and bridge in Sava (chl-a distribution Mishra and Mishra NDCI algorithm, 26-9-2021).

Figure 36 Locations of a ship and bridge in Sava (pan, 26-9-2021).

Figure 37 Elevated chl-a concentrations at locations of ships in Danube (chl-a distribution Mishra and Mishra NDCI algorithm, 26-9-2021).

Figure 38 Locations of ships in Danube (pan, 26-9-2021).

Figure 39 Chl-a distibution at confluence (chl-a distribution Mishra and Mishra NDCI algorithm, 26-9-2021).

Figure 40 Confluence of Danube and Sava (pan, 26-9-2021).

45 4.1.3. Validation of the original algorithms

A small but well in time with satellite overpass matching dataset comprising of 11 in-situ samples was used to validate the original band-ratio and NDCI algorithms for application on the research area. This dataset is smaller than those used in the USA when developing the models. Gurlin et al. (2011) used two separate datasets for the calibration and validation of their band-ratio model. 89 measurements gathered at the Fremont Lakes Stage Recreation Area in Nebraska, USA in 2008 were used for the calibration of their model. The validation took place with a dataset comprising of 63 measurements from that same area, taken in 2009. The Mishra and Mishra NDCI algorithm was calibrated and validated using both a simulated dataset and a field dataset (Mishra & Mishra, 2012). For the calibration and validation using the field dataset, of the 49 samples, 29 were used for calibration and 20 for validation. The dataset for validation for this research is smaller than those of others, but with 11 samples a validation can still be performed of which the results should be a good initial indication of applicability of chl-a concentration algorithms on rivers.

As discussed, the dataset used in this research is, apart from site 12, homogenous for an area where the chl-a concentrations vary more. Results of several algorithms do show that the chl-a concentration is higher in the Danube upstream of the confluence than at most of the sample locations. The results of the Mishra and Mishra NDCI algorithm show the chl-a concentration ranges from 16μg/L to 20μg/L.

Even though the concentrations measured range from 2.1μg/L to 7.0μg/L and do not cover the full range of concentrations present in the research area, it was assumed they would have enough variance for an indicative validation to be performed with.

4.1.4. Calibration of the algorithms

The calibration was performed using a quadratic regression. Another form of regression, like a linear regression, was not chosen because the original algorithms from Gurlin et al. (2011) and Mishra and Mishra (2012) used a quadratic regression in their most successful calibrations. Performing the same regression would offer the best comparison between the original and calibrated algorithms. The band-ratio model was calibrated using a linear regression in Gurlin et al. (2011) as well, but this algorithm had a lower performance.

In Gurlin et al. (2011) and Mishra and Mishra (2012), calibration took place with a dataset comprising of respectively 89 and 29 samples. The full dataset for this research contained only 11 samples and there were no options to extend this dataset for this research. To still work with this limited dataset, the question arose if this dataset was enough for an accurate calibration. 11 samples were not enough to calibrate a model to depict accurate absolute values for chl-a concentrations. It was assumed that 11 would be enough to calibrate a model with the purpose of making estimations of chl-a concentrations. The choice was made to use all data for the calibration of the algorithm, leaving none for a separate validation. With only 11 samples, it was assumed that every sample was needed for the most accurate calibration. Reserving samples for a separate validation would decrease the already limited accuracy.

4.1.5. Validation of the calibrated algorithms

Gurlin et al. (2011) used two datasets for calibration and validation, while Mishra and Mishra (2012) reserved part of one dataset for validation. The choice was made not to do either in this research and the algorithms were calibrated using the full dataset without reserving data for validation. To make a more reliable validation, the algorithms were recalibrated with only 8 samples, reserving 3 for validation. Validating the calibrated algorithms using the same data used for training yielded bias results that did not show an accurate performance of the algorithms. The performance of the calibrated band-ratio algorithm was presented as outstanding after validation, while the recalibration

46 showed a much poorer performance. Also, the resulting chl-a distribution map of the calibrated band-ratio algorithm did not show the spatial distributions that were expected and visible in the results from the other 3 algorithms.

The recalibration and validation of the algorithms did put the performance in perspective. The resulting performance measures is not the performance of the calibrated algorithms using the full dataset, as samples are missing from the dataset. It does however serve as an indication, and even a minimum value of the performance, as it can be assumed that accuracy and performance of the algorithm increases with a more extensive dataset.