GSA Annual Meeting in Phoenix, Arizona, USA - 2019

Paper No. 111-8
Presentation Time: 9:00 AM-6:30 PM


PEREIRA, O.J.R.1, MERINO, E.R.1, MONTES, C.R.2, BARBIERO, Laurent3, LUCAS, Yves4, REZENDE-FILHO, Ary T.5 and MELFI, Adolpho J.1, (1)University of São Paulo - USP, Institute of Energy and Environment - IEE-USP, São Paulo, Brazil, (2)University of São Paulo - USP, CENA, NUPEGEL, Piracicaba, Brazil, (3)GET, IRD, CNRS, UPS, OMP Toulouse, France, (4)Université de Toulon, Provence, France, (5)Universidade Federal do Mato Grosso do Sul, Campo Grande, Brazil

Google Earth Engine (GEE) is the largest satellite imagery cloud-based platform. This tool has increased the feasibility of remote sensing-based environmental analysis from local to planetary scales. Among Earth systems, wetlands are widely recognized for their vital role. Yet, depending on the temperature, pH, and redox conditions, they can be important sinks or sources of greenhouses gases. The Nhecolândia region (southern Pantanal wetland), is a unique lacustrine system with more than 10,000 lakes and ponds with saline and freshwater coexisting in close proximity.

Here, we present the first model for predicting water pH values in the Nhecolândia lakes based on Landsat TM/ETM+ and OLI images time-series (2002 – 2017) using GEE. We used the TOA Reflectance collection of Landsat TM5, ETM+ 7, and OLI 8 and cloud-free pixels. We obtained 2081 surface reflectance scenes and merged them into a single collection. Additionally, we included synthetic band indexes: (a) Normalized Difference Vegetation Index (NDVI), (b) Automated Water Extraction Index with no shadow (AWEInsh), (c) Normalized Difference Water Index (NDWI), and (d) Modified Normalized Difference Water Index (MNDWI). We presented a field-validated analysis aiming to predict the pH values of the Nhecolândia lakes; analyze spectral signature variations of the lakes with changes in pH; correlate the different types of lakes with relief landforms and drainage network. The prediction of pH was made by testing seasonal filters, considering linear and non-linear regression methods, with field-measured pH values as dependent variables and median Landsat bands as explanatory variables. We applied linear multiple regression and symbolic regression based on Genetic Programming (GP).

The regression model presented an r² value of 0.81 and pH values ranged from 4.69 to 11.64. The highest correlation considered median pixel values of the high water season Landsat collection from 2002 to 2017, with an r² value of 0.72 and an RMSE of 0.90. We mapped 12,150 lakes, with saline lakes accounting for 7.25% of them. The GEE platform was a key factor in the feasibility of this study, allowing us to create a new pathway to analyze the salinity of inland water bodies, given that most of the models and products mainly focus on ocean waters or on deltaic environments.