Flood Susceptibility Modelling Using Remote Sensing – Machine Learning Approach and Optical Water Quality Analysis of Vembanad Lake System In Kerala, India

Abstract

Wetlands are essential ecosystems that play a significant role in mitigating the impacts of climate change. Wetlands store large amounts of carbon and help to regulate the climate by reducing the amount of carbon dioxide in the atmosphere. They also help to reduce the impacts of extreme weather events, such as floods and hurricanes, by absorbing and retaining water. However, wetlands are also vulnerable to the effects of natural and anthropogenic factors, which can alter their hydrology and lead to the loss of wetland habitats. It is crucial to protect and preserve wetlands to maintain their vital role in mitigating the impacts of climate change. The wetland functions, commodities, and services are lost due to upland land use activities. Hence, accurate and up-to-date information on the upland regions around wetlands is essential. The present research considers the Vembanad Lake System (VLS) in Kerala, India, which is specifically affected by challenging issues to its health and survival. The study area faces threats like encroachment and climate change resulting in floods and alteration in the precipitation patterns. Further, the lake system is endangered by the deteriorating quality of incoming water. Thus, the overall spatio-temporal analysis is critical in protecting and managing water resources in the study region. Anthropogenic activities result in a massive Land Use and Land Cover (LULC) change, and it has become a prominent issue for decision planners and conservationists due to inappropriate growth and its effect on natural ecosystems. As a result, the change in LULC for the short term, i.e., within a decade, is carried out using three Machine Learning (ML) approaches, Random Forest (RF), Classification And Regression Trees (CART), and Support Vector Machine (SVM), on the Google Earth Engine (GEE) platform. When comparing the three techniques, SVM performed poorly at an average accuracy of around 82.5%, CART being the next at 87.5%, and the RF model being good at an average of 89.5%. The RF outperformed the SVM and CART in almost identical spectral classes, such as barren land and built-up areas. As a result, RF- classified LULC is considered to predict the Spatio-temporal distribution of LULC transition analysis for 2035 and 2050. This analysis was conducted in Idrisi TerrSet software using the Cellular Automata (CA) - Markov chain analysis. The model's efficiency is evaluated by comparing the projected 2019 image to the actual 2019 iclassified image. The model efficiency obtained was good, with more than 94.5% accuracy for the classes except for barren land, which might have resulted from the recent natural calamities and the accelerated anthropogenic activity in the study area. Floods have claimed the lives of countless people and caused significant property damage, putting their livelihoods in jeopardy. The study area faced adverse mishappening during the 2018, 2019, and 2021 floods due to the torrential rainfall events. Estimations of flood-inundated areas are prepared from 2018, 2035, and 2050 LULC maps. The extent of flood inundation during the 2018 floods and the possible flood inundation region for the projected LULC in 2035 and 2050 are determined. From the analysis of the 2018 classified image, 14.7 km2 of built-up area was found inundated during the year 2018 floods. The scenario of the 2018 flood event is used to quantify the flood that may occur and inundate the projected LULC 2035 and 2050 scenarios. It is found that the flood will affect about 19.87 km2 and 23.32 km2 of the built-up region, majorly for the 2035 and 2050 projected scenarios, respectively. The goal of this research is to construct effective decision tree-based ML models such as Adaptive Boosting (AdaBoost), RF, Gradient Boosting Machines (GBM), and Extreme Gradient Boosting (XGBoost) for integrating data, processing and generating flood susceptibility maps. Eighteen conditioning parameters, including seven categorical and eleven numerical data, are used for flood modelling using ML. These seven categorical data are converted into 50 numerical data, resulting in a total input data of 61. The Recursive Feature Elimination (RFE) is utilized as the feature selection technique, and 22 layers are chosen to feed into the ML models to generate the flood susceptibility maps. The efficiencies of the models are evaluated using Receiver Operating Characteristic – Area Under Curve (ROC-AUC), F1 score, Accuracy, and Kappa. According to the results obtained, all four ML models demonstrated fairly good performance. However, XGBoost fared well in terms of the model's metrics. The ROC-AUC values of XGBoost, GBM, and AdaBoost for the testing dataset are 0.90, whereas 0.89 for RF. The accuracy varied significantly among the four models, with XGBoost scoring 0.92, followed by GBM (0.88), RF (0.87), and AdaBoost (0.87). The resulting flood susceptibility map can be utilized for early mitigation actions during future floods and for land use planners and emergency managers, assisting in reducing flood risk in regions prone to this hazard. iiWater quality is one of the essential parameters of environmental monitoring; even a slight variation in its characteristics may significantly influence the ecosystem. The water quality of Vembanad Lake is affected by anthropogenic effects such as industrial effluents and tourism. The optical parameters representing water quality, such as diffuse attenuation (Kd), turbidity, Suspended Particulate Matter (SPM), and Chlorophyll-a (Chl-a), are considered in this study to evaluate the water quality of the Vembanad Lake. As this lake is regarded as of ecological importance by the Ramsar Convention and has faced severe concerns over recent years, there was a substantial change in the water quality during the lockdowns of the COVID-19 pandemic. This research aimed to examine the change in water quality using optical data from Sentinel-2 satellites in the ACOLITE processing software from 2016 to 2021. The analyses showed a 2.5% decrease in the values of Kd, whereas SPM and turbidity show a reduction of about 4.3% from the year 2016 to 2021. The flood and the COVID lockdown had an impact on the improvement in the quality of water from 2018 to 2021. The findings indicated that the reduction in industrial activities and tourism had a more significant effect on the improvement in the water quality of the lake. There was no substantial change in the Chl-a until 2020, whereas an average decrease of 12% in Chl-a values was observed throughout 2021. This decrease can be attributed to the reduction in the lake's Hydrological Residence Time (HRT). The outcome of this research depicts augmentation of the change in the LULC pattern and its prediction, future flood-inundation regions, flood susceptibility mapping, and the lake's water quality. The findings of this research work will be a valuable reference to help the government and Non-Government Organisations (NGOs) during strategic planning.