Hyperspectral retrieval of phycocyanin in potable water sources using genetic algorithm–partial least squares (GA–PLS) modeling

• Published in: International journal of applied earth observation and geoinformation 2012-08, Vol.18, p.368-385
• Main Authors: Song, Kaishan, Li, Lin, Li, Shuai, Tedesco, Lenore, Hall, Bob, Li, Zuchuan
• Format: Article
• Language: English
• Subjects: AISA; Applied geophysics; Cyanobacteria; Earth sciences; Earth, ocean, space; Exact sciences and technology; GA–PLS; Hyperspectral; Internal geophysics; Phycocyanin; Semi-empirical model
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2012.03.013

▸ We estimate cyanobacteria pigment phycocyanin (C-PC) in water supply sources. ▸ A hybrid GA–PLS model is developed for remotely estimating PC concentration. ▸ Three-band model shows stable performance for C-PC estimation, but spectral band position affected by optically active constituents. ▸ Hyperspectral remote sensing coupled GA–PLS modeling provides useful tool for cyanobacteria monitoring. Eagle Creek, Morse and Geist reservoirs, drinking water supply sources for the Indianapolis, Indiana, USA metropolitan region, are experiencing nuisance cyanobacterial blooms. Hyperspectral remote sensing has been proven to be an effective tool for phycocyanin (C-PC) concentration retrieval, a proxy pigment unique to cyanobacteria in freshwater ecosystems. An adaptive model based on genetic algorithm and partial least squares (GA–PLS), together with three-band algorithm (TBA) and other band ratio algorithms were applied to hyperspectral data acquired from in situ (ASD spectrometer) and airborne (AISA sensor) platforms. The results indicated that GA–PLS achieved high correlation between measured and estimated C-PC for GR (RMSE=16.3μg/L, RMSE%=18.2; range (R): 2.6–185.1μg/L), MR (RMSE=8.7μg/L, RMSE%=15.6; R: 3.3–371.0μg/L) and ECR (RMSE=19.3μg/L, RMSE%=26.4; R: 0.7–245.0μg/L) for the in situ datasets. TBA also performed well compared to other band ratio algorithms due to its optimal band tuning process and the reduction of backscattering effects through the third band. GA–PLS (GR: RMSE=24.1μg/L, RMSE%=25.2, R: 25.2–185.1μg/L; MR: RMSE=15.7μg/L, RMSE%=37.4, R: 2.0–135.1μg/L) and TBA (GR: RMSE=28.3μg/L, RMSE%=30.1; MR: RMSE=17.7μg/L, RMSE%=41.9) methods results in somewhat lower accuracy using AISA imagery data, which is likely due to atmospheric correction or radiometric resolution. GA–PLS (TBA) obtained an RMSE of 24.82μg/L (35.8μg/L), and RMSE% of 31.24 (43.5) between measured and estimated C-PC for aggregated datasets. C-PC maps were generated through GA–PLS using AISA imagery data. The C-PC concentration had an average value of 67.31±44.23μg/L in MR with a large range of concentration, while the GR had a higher average value 103.17±33.45μg/L.