Machine Learning Algorithms for Chromophoric Dissolved Organic Matter (CDOM) Estimation Based on Landsat 8 Images

Algal biomass mapping of eutrophic lakes using a machine learning approach with MODIS images

Algal blooms are a widespread issue in eutrophic lakes. Compared with the satellite-derived surface algal bloom area and chlorophyll-a (Chla) concentration, algae biomass is a more stable way to reflect water quality. Although satellite data have been adopted to observe the water column integrated algal biomass, the previous methods mostly are empirical algorithms, which are not stable enough for widespread use. This paper proposed a machine learning algorithm based on Moderate Resolution Imaging Spectrometer (MODIS) data to estimate the algal biomass, which was successfully applied to a eutrophic lake in China, Lake Taihu. This algorithm was developed by linking Rayleigh-corrected reflectance to in situ algae biomass data in Lake Taihu (n = 140), and the different mainstream machine learning (ML) methods were compared and validated. The partial least squares regression (PLSR) (R² = 0.67, mean absolute percentage error (MAPE) = 38.88 %) and support vector machines (SVM) (R² = 0.46, MAPE = 52.02 %) performed poor satisfactory. In contrast, random forest (RF) and extremely gradient boosting tree (XGBoost) algorithms had higher accuracy (RF: R² = 0.85, MAPE = 22.68 %; XGBoost: R² = 0.83, MAPE = 24.06 %), demonstrating greater application potential in algal biomass estimation. Field biomass data were further used to estimate the RF algorithm, which showed acceptable precision (R² = 0.86, MAPE < 7 mg Chla). Subsequently, sensitivity analysis showed that the RF algorithm was not sensitive to high suspension and thickness of aerosols (rate of change <2 %), and inter-day and consecutive days verification showed stability (rate of change <5 %). The algorithm was also extended to Lake Chaohu (R² = 0.93, MAPE = 18.42 %), demonstrating its potential in other eutrophic lakes. This study for algae biomass estimation provides technical means with higher accuracy and greater universality for the management of eutrophic lakes.

Remote sensing of fluorescent humification levels and its potential environmental linkages in lakes across China

The pollution or eutrophication affected by dissolved organic matter (DOM) composition and sources of inland waters had attracted concerns from the public and government in China. Combined with remote sensing techniques, the fluorescent DOM (FDOM) parameters accounted for the important part of optical constituent as chromophoric dissolved organic matter (CDOM) was a useful tool to trace relative DOM sources and assess the trophic states for large-scale regions comprehensively and timely. Here, the objective of this research is to calibrate and validate a general model based on Landsat 8 OLI product embedded in Google Earth Engine (GEE) for deriving humification index (HIX) based on EEMs in lakes across China. The Landsat surface reflectance was matched with 1150 pairs fieldtrip samples and the nine sensitive spectral variables with good correlation with HIX were selected as the inputs in machine learning methods. The calibration of XGBoost model (R² = 0.86, RMSE = 0.29) outperformed other models. Our results indicated that the entire dataset of HIX has a strong association with Landsat reflectance, yielding low root mean square error between measured and predicted HIX (R² = 0.81, RMSE = 0.42) for lakes in China. Finally, the optimal XGBoost model was used to calculate the spatial distribution of HIX of 2015 and 2020 in typical lakes selected from the Report on the State of the Ecology and Environment in China. The significant decreasing of HIX from 2015 to 2020 with trophic states showed positive control of humification level of lakes based on the published document of Action plan for prevention and control of water pollution in 2015 of China. The calibrated model would greatly facilitate FDOM monitoring in lakes, and provide indicators for relative DOM sources to evaluate the impact of water protection measures or human disturbance effect from Covid-19 lockdown, and offer the government supervision to improve the water quality management for lake ecosystems.

Estimation of dissolved organic carbon from inland waters at a large scale using satellite data and machine learning methods

Dissolved Organic Carbon (DOC) in inland waters plays an essential role in the global carbon cycle and has significant public health effects. Machine learning (ML) together with remote sensing has emerged as a powerful and promising combination to quantify water quality parameters from space. However, inland water sample data for DOC is limited. Hence, little is known about the potential to quantify DOC content in inland waters, especially over large-scale areas. This study presents the first attempt to estimate DOC in inland waters over a large-scale area using satellite data and ML methods with the newly published open-source dataset AquaSat. Four ML approaches, namely Random Forest Regression (RFR), Support Vector Regression (SVR), Gaussian Process Regression (GPR), and a Multilayer Backpropagation Neural Network (MBPNN) were trained using more than 16 thousand samples across the continental United States matched with satellite data from Landsat 5, 7 and 8 missions. Satellite data from the Landsat missions were further extended with environmental data from the ERA5-Land product and used as input to train the ML algorithms. Our results show that including environmental data as inputs considerably improved the prediction of DOC for all ML algorithms, with GPR showing the most promising performance results with moderate estimation errors (RMSE: 4.08 mg/L). Permutation feature importance analysis showed that the wavelength range in the visible Green band (from Landsat) and the monthly average air temperature (from ERA5-Land) were the most important variables for the ML approaches. The results demonstrate the predictive strength of GPR and its useful feature to derive per pixel standard deviations for detailed analysis. Our results further highlight the important role of considering environmental processes to explain DOC variations over large scales. The application and performance of the GPR in mapping spatiotemporal variations of DOC in an entire water body were discussed by taking Lake Okeechobee (the 8^th largest freshwater lake in the U.S.) as an illustrative example. While performance evaluation showed that DOC concentrations can be retrieved with adequate accuracy, algorithm development was challenged by the heterogenous nature of large-scale open source in situ data, issues related to atmospheric correction, and the low spatial and temporal resolution of the environmental predictors. This research demonstrates how open source, large-scale datasets like AquaSat in combination with ML and satellite remote sensing can make research toward large-scale estimation of inland water DOC more realistic while highlighting its remaining limitations and challenges.

Estimation of chromophoric dissolved organic matter and its controlling factors in Beaufort Sea using mixture density network and Sentinel-3 data

With the warming of the high-latitude regional climate, melting of permafrost, and acceleration of hydrological cycles, the Arctic Ocean (AO) has undergone a series of rapid changes in the past decades. As a dominant optical component of the AO, the variations in chromophoric dissolved organic matter (CDOM) concentration affect the physiological state marine organisms. In this study, machine learning retrieval model based on in situ data and mixture density network (MDN) was developed. Compared to other models, MDN model performed better on test data (R² = 0.83, and root mean squared error = 0.22 m^-1) and was applied to Sentinel-3 OLCI data. Afterward, the spatiotemporal distribution of CDOM during the ice-free (June-September) from 2016 to 2020 in the Beaufort Sea was obtained. CDOM concentration generally exhibited an upward trend. The maximum monthly average CDOM concentration appeared in June and gradually decreased thereafter, reaching its lowest value in September of each year. The maximum value appeared in June 2020 (0.91 m^-1), and the minimum value was observed in September 2017 (0.81 m^-1). The CDOM concentration nearshore was higher than that in other areas; and gradually decreased from offshore to the open sea. CDOM was highly correlated with salinity (R² = 0.49) and discharge (R² = 0.53), and the tight correlation between salinity and CDOM further suggested that terrestrial inputs were the main source of CDOM in the Beaufort Sea. However, sea level pressure contributed to the spatial variations in CDOM. When southerly wind prevailed and wind direction was aligned with the CDOM diffusion direction, the wind accelerated the diffusion of CDOM into the open sea. Meanwhile, seawater was diluted by the sea ice melting, resulting in decrease in CDOM concentration. Herein, this paper proposed a robust and near real-time method for CDOM monitoring and influence factor analysis, which would promote the understanding of AO CDOM budgets.

Estimation of Chlorophyll-a Concentrations in Small Water Bodies: Comparison of Fused Gaofen-6 and Sentinel-2 Sensors

Chlorophyll-a concentrations in water bodies are one of the most important environmental evaluation indicators in monitoring the water environment. Small water bodies include headwater streams, springs, ditches, flushes, small lakes, and ponds, which represent important freshwater resources. However, the relatively narrow and fragmented nature of small water bodies makes it difficult to monitor chlorophyll-a via medium-resolution remote sensing. In the present study, we first fused Gaofen-6 (a new Chinese satellite) images to obtain 2 m resolution images with 8 bands, which was approved as a good data source for Chlorophyll-a monitoring in small water bodies as Sentinel-2. Further, we compared five semi-empirical and four machine learning models to estimate chlorophyll-a concentrations via simulated reflectance using fused Gaofen-6 and Sentinel-2 spectral response function. The results showed that the extreme gradient boosting tree model (one of the machine learning models) is the most accurate. The mean relative error (MRE) was 9.03%, and the root-mean-square error (RMSE) was 4.5 mg/m3 for the Sentinel-2 sensor, while for the fused Gaofen-6 image, MRE was 6.73%, and RMSE was 3.26 mg/m3. Thus, both fused Gaofen-6 and Sentinel-2 could estimate the chlorophyll-a concentrations in small water bodies. Since the fused Gaofen-6 exhibited a higher spatial resolution and Sentinel-2 exhibited a higher temporal resolution.