What water color parameters could be mapped using MODIS land reflectance products: A global evaluation over coastal and inland waters

MODIS surface reflectance product (R_land) has been used to monitor waters due to its free availability and higher spatial resolution than MODIS ocean bands. However, its applicability in aquatic remote sensing has not been sufficiently assessed. Some fundamental questions such as the following need to be addressed: How does the R_land product perform in global inland and coastal waters? What water color parameters can be mapped using R_land? This study provided a comprehensive evaluation of the performance of MODIS R_land products against a field optical dataset containing 4143 reflectance spectra, 2320 chlorophyll-a (Chla) samples, and 1467 suspended particulate matter (SPM) samples across global nearshore coastal and inland waters. The results showed that R_land significantly overestimated remote sensing reflectance, particularly in the bands of 469 nm and 859 nm. The noticeable negative values and patchiness were found in the R_land imagery, and existing algorithms did not estimate satisfactory Chla and SPM from R_land across the global inland and coastal waters. Furthermore, we tested popular machine-learning approaches, such as random forest (RF), support vector machine, XGBoost, and deep neural networks, to examine the potential of the R_land product in estimating SPM and Chla. Machine learning models were found to outperform the state-of-the-art algorithms for SPM retrievals from R_land. Specifically, RF and XGBoost showed the good performance with mean absolute errors of -25.0% and mean absolute percentage error of -23% for a broad SPM range of 10-500 mg L-1. Yet, machine learning models cannot retrieve reliable Chla from R_land with approximately 55% uncertainty due to the limited spectral information and uncertainty of R_land products. This implicated that R_land might be able to quantify the parameters that are closely related to SPM (e.g., water clarity and extinction coefficients) in most waters; however, it is challenging to quantify pigments like Chla in waters from R_land. We conclude that R_land might not be an optimal data source for monitoring inland and coastal waters, despite the ease of using this product and its higher spatial resolution than the MODIS ocean bands.

Automated summary

This study comprehensively evaluated the performance of MODIS R_land products in global inland and coastal waters and found that it overestimates reflectance and cannot accurately estimate chlorophyll-a and suspended particulate matter. Machine learning models showed good performance in estimating suspended particulate matter but unreliable in estimating chlorophyll-a.