• No results found

Chapter 1: Evaluation and optimization of remote sensing techniques for detection of

1.6. Conclusion and recommendations

37

The use of FLH for red tide detection and tracking over time (Hu et al., 2005) has been shown useful for the WFS. A concern when using only FLH is that sediment resuspension events will lead to overestimating FLH and yield false positives. Excellent results can be obtained through the combined use of FLH and Rrs(555) to decrease the likelihood of resuspension events as being falsely detected as red tides.

1.6. Conclusion and recommendations

The main objective of this chapter was to evaluate the performance of several published K. brevis remote sensing detection techniques (1-Stumpf CHL anomaly (Stumpf et al., 2003), 2-Tomlinson spectral shape (2-Tomlinson et al., 2009), 3- Cannizzaro bbp ratio (Cannizzaro et al., 2008; 2009; Hu et al., 2011), 4- Amin RBD-KBBI technique (Amin et al., 2009), 5- a technique similar to the one suggested by Carvalho et al. (2010) using Rrs (555), 6- a combination of techniques (hybrid technique) as suggested by Carvalho et al., 2010) in different regions of the WFS and to improve their performance by fine-tuning their criteria. A second objective was to test the performance of a new and simpler detection technique using only FLH and Rrs (555) satellite data. An extensive dataset of 2323 in situ and MODIS data matching pairs were used to

systematically optimize variables and coefficients used in each of these published HAB detection methods. Statistical measures commonly used for image analysis techniques (e.g., FM,

specificity and sensitivity) were used to optimize the techniques by finding the criteria and

thresholds that result in the best performance. Before the optimization, the average FM was 0.47, ranging from 0.34 (Stumpf CHL anomaly) to 0.57 (Amin RBD-KBBI). Optimization increased the FM for all techniques from an average of ~0.47 to ~0.59. The percentage of false positives was reduced by almost 50% for most of the techniques. The addition of a FLH criterion improved the performance of each method, and in particular that using the spectral shape at 490 nm and backscattering coefficients. Maximum FM was obtained by the optimized RBD-KBBI (0.63). The new practical method, which identifies K. brevis blooms as those pixels showing FLH above 0.033 mW cm-2 µm-1 sr-1 and Rrs (555) below 0.007 sr-1, performed similar to the more complex RBD-KBBI technique. This new, simple approach yielded an FM of 0.62 and generated only 3%

false negatives

.

The Northern region and Southern regions of Florida represented a challenge for most of the techniques due to the high absorption by CDOM in the North or high reflectance in shallow waters in the South. The Amin RBD-KBBI technique and the new Rrs-FLH technique seemed to perform better in these regions as well as in the Central WFS. FLH alone is an excellent satellite product for the detection of blooms, and during the optimization process FLH was the satellite product helping to reduce false positives, especially those due to dark waters associated with high absorption of CDOM in riverine waters. The use of Rrs (555) helped reduce the problems in regions with high bottom reflectance or resuspended sediment concentration. The combination of FLH-Rrs is in theory from the same concept as the RBD-KBBI technique, however it is a much simpler technique, easier to implement with existing satellite data products, and it allows for easier adjusting of the thresholds (to increase sensitivity or specificity) based on the specific needs of different user groups.

All detection techniques proved to be useful for the detection of HABs in the WFS. The optimization exercise allowed for improving each individual technique. The extensive evaluation effort led to the following recommendations to improve their performance and to simplify the implementation of operational HAB detection algorithms:

1) The chlorophyll anomaly cannot be fully evaluated here as an automated technique because it includes a heuristic model that requires ancillary environmental observations and a visual inspection by an image analyst. The heuristic model has been adapted to the WFS based on prior knowledge of the region by the author. The implementation of the heuristic model in other parts of the GOM will require similar knowledge of the dynamics of regional phytoplankton communities, of optical properties of the blooms, and also of the regional physical oceanography. The performance of this technique without the heuristic model was poor in comparison with the other techniques.

2) Including Rrs(555) in HAB detection algorithms increased sensitivity, and including FLH increases accuracy. The differences in methods to estimate bbp(555) from satellite data introduces errors and variability in HAB detection techniques. Rrs (555) image products are readily available from NASA as a standard ocean color product. Without additional

 

39

corrections, simply using Rrs (555) leads to higher numbers of false positives. I recommend also using a FLH threshold to remove pixels with high uncertainty. The optimization results suggested a threshold of FLH between 0.02-0.04 mW cm-2 µm-1 sr-1, with best results using 0.033 mW cm-2 µm-1 sr-1.

3) The spectral shape technique showed high sensitivity and lower specificity. The visual analysis also showed a high tendency for false positives especially in the Northern region of Florida. I recommend including FLH and specifically a FLH threshold between 0.02-0.04 mW cm-2 µm-1 sr-1.

4) The RBD-KBBI technique had the best performance visually and numerically, compared to all other techniques; it is specific although less sensitive. The modified version of the RBD-KBBI (KBBI<0.8*RBD) greatly improved the sensitivity of the technique.

5) While FLH has proven to be a very useful measure to improve HAB detection techniques, at present none of the US ocean-observing satellite missions planned for launch before about 2020 include the FLH bands. In particular, the most recently launched Visible Infrared Imaging Radiometer Suite (VIIRS, 2011 – present) is not equipped with FLH bands. I strongly recommend that operational agencies responsible for monitoring HABs nationally and internationally include the FLH capabilities in future sensors.

Chapter 2: Visualization and Quantification of Harmful Algal