I’m trying to retrieve chl_a concentration using C2RCC in my study area.
I tried to process the image using default processing parameters vs with adjusted salinity, temperature, elevation, and ozone suitable to my study area with the concept of adjusting processing parameters values can improve the results (as also stated here by @ABraun Retriving chlorophyll and turbidity from sentinel data - #57 by ABraun )
But after extracting the pixel value from the model and comparing it to the in-situ data, the default one has a slightly higher R2 value (R2=0.08 for default model, and R2=0.02 for adjusted model).
Moreover, both my C2RCC model has chl_a concentration with the range value 0.01-0.5 mg/m3 while my in-situ data has chl-a range about 2-10 mg/m3.
I also tried to calculate the chl_factor and chl_exponen using my in-situ data (like what’s being recommended before by @abruescas and @marpet ) but seems like it didn’t work and the model gave con_chl with even lower pixel value.
Would you please enlighten me is there any other factor I have to consider to improve my C2RCC model?
Your lab data essentially has a cluster of values that are low for both lab and pixel extraction, and just two outliers where pixel extraction and lab results are contradictory.
Outliers have a big influence on regressions, so some explanations is needed.
A log-log transform may be more appropriate, but you should examine the outliers carefully. If parts of your your study area are shallow or close to land you should consider distance from land and/or water depth. Optical sensors have a Gaussian like response surface, but a very bright area away from the pixel center (e.g., shallow water with sandy) bottom can produce a bright pixel.
Thank you for helping me out, sir.
I’ve tried to remove the outliers but it results in even lower R2
is it possible to increase my model chl-a concentration range besides working on my lab data?
also, I’ve tried using natural-log for my data by converting the chl-a value from pixel extraction and lab to natural-log and making the regression then applying the equation to my C2RCC model using band math, am I doing the correct step?
It is vey difficult to obtain a good correlation with low number of matchups. If you are using Pixel Extraction, I would recommend to extract at least a 3 x 3 macropixel over your in situ measure. Usually there is a lag between the in situ and the satellite image, of hours or even days, and water is a dynamic environment, If yoy extract a 3 x 3 macopixel, you can calculate the average value and try to compare again, but you should also avoid introducng uncertainties. Read carefully Bailey and Werdell before using the average: * 10.1016/j.rse.2006.01.015
Thank you for your answer, Mrs. @abruescas
I’ve tried to extract my model using 3x3 windows and even 5x5 but the result doesn’t make any big difference, however i also agree that my limited sample data is causing the low correlation.
and thank you so much for your paper recommendation, have a great day ma’am