GEE-based Classification Of African Wetlands

Wetlands constitute one of the earth's most vital ecosystems.This is because they offer many benefits to society which include among them water purification,flood risk reduction,water and soil conservation,protection of shorelines,aesthetic and recreational values.As a result of their invaluable benefits,they have become to be known as the"kidneys of the earth".Despite their well-known values,wetlands still face the danger of being drained and being replaced with other human land use types such as agriculture and settlements.Spatially-explicit information or knowledge of the wetlands is key to preserving them in the face of land-use changes.In Africa spatial information on wetlands is insufficient for it to be used in continental wetland management.The chief aim of this study was to leverage contemporary cloud computing platforms such as Google Earth Engine(GEE)to classify and map the wetlands of Africa from 2001 to 2017 using Landsat data.Consequently,the study set its objectives as follows;(i)to develop an algorithm for automatic classification of wetlands from Landsat satellite imagery using GEE Application Programming Interface(API).(ii)to use GEE to automatically classify and delineate Africa's wetlands using Landsat satellite imagery.(iii)to determine how the wetland areas,varied during the wet and dry seasons from year 2001 to 2017 and what factors influenced this variation.(iv)to create a GEE-based web application of map displaying Africa‘s wetlands,their types and distribution from the automatic classification results of the developed algorithm.(v)to use Landsat imagery to classify wetlands using visual observation.In order to achieve these objectives,two approaches were adopted namely an automatic and visual based classification methods.For the automatic classification method,a remote sensing based automatic wetland classification system was developed.The MODIS MCD12Q1 Landcover Type 2 dataset was first downscaled and used to train the Landsat 7 composites using random forest classification method for the wet and dry seasons of each year from2001 to 2017.The classification accuracy for each period was greater than 0.75.The study adopted Uganda's national definition of wetlands which defines them as landcovers which are permanently or seasonally saturated with water.The Modified Normalized Difference Water Index(MNDWI)was used to determine water saturated landcovers from the classified images and produce different wetland classes.8 wetland classes namely Water Bodies(Lakes,Reservoirs,Rivers),Forested Swamp,Shrub-dominated Wetland,Woody Savanna Wetland,Marshland,Permanent Inundated Wetland,Cropland-dominated Wetland and Pan/Saline Wetlands were derived and determined from the classified Landsat landcover image.The areal results indicated that averagely,the African wetland extent was approximately 1190 000km~2 during the wet season and 780 000km~2 during the dry season from 2001 to 2017.As a percentage of the total African surface area wetlands covered 3.92%and 2.58%during the wet and dry seasons respectively.Forested swamps,water bodies(lakes,rivers and reservoirs)and marshlands made up about 75%of the wetland area.The remaining 25%of the wetland area was made up of shrub-dominated,cropland-dominated,pan or saline,woody savanna and permanent inundated wetlands.The largest concentration of wetlands occurs in the wetlands of the major river and lake systems of Lakes Victoria in Uganda and Tanzania,Tanganyika in Tanzania and D.R.C,Turkana in Kenya and Zambezi River in the southern,central and eastern subregions.In the northern and western subregions,a large occurrence of wetlands is found in Egypt along Lake Nasser-Nile River confluence,Lake Chad,Sudan,Niger and Mali.The countries with the largest occurrence of wetlands were found to be the D.R.C and Tanzania.The islands of Comoros,Mauritius,Seychelles and Cape Verde have the least wetlands occurrence in Africa.The study also determined that Africa's wetlands are largely concentrated along the four major river ecosystems of Niger,Nile,D.R.C and Zambezi.With regards to the variation of area of wetland coverage in Africa for the period under review,the study determined that there is no constant pattern of area change.The area coverage of wetlands increased or decreased from one year to another.Also,the area coverages of specific wetland classes varied from each year as their areas either increased or decreased with seasons and different years.The study also aimed to create a GEE-based application that is capable of automatically classifying Africa's wetlands and display the results of the classification as a map.As a result,a publicly accessible application leveraging on GEE's computing capabilities was developed.The web-based automatic wetland classification application was developed using the Earth Engine Apps platform.A simple graphical user interface(GUI)whose elemental widgets was made up of textboxes to collect input data,buttons to perform command operations,labels to display static information and panels to act as containers for other widgets and overall layout of the user interface was designed.The developed web application can be accessed on this URL:https://mikemurefu.users.earthengine.app/view/wetland-classification-app.The developed web application can classify Africa's wetlands according to the automatic landcover derived classification system set out in this study.A visual validation of the classification results from the web application using satellite images of well-known African wetlands proved that the application is suitable to classify wetlands in Africa.This study also classified wetlands using visual observation method to establish the training samples.Wetland classes including Lakes/Rivers/Ponds,Marsh/Swamps and Pan/Saline Wetlands were determined and classified using three supervised classification algorithms namely Classification and Regression Trees(CART),Random Forests and Support Vector Machine(SVM)for the years 2001,2003,2006,2009,2011 and 2014.The CART algorithm produced the highest classification accuracies followed by Random Forest and finally the SVM algorithm.The average classification accuracy of CART was 0.998.For the random forests and SVM algorithms,their average classification accuracies were 0.958 and 0.881 respectively.The area coverage of the Lakes/Rivers/Ponds wetland class was almost equal in all the classification algorithms but differed for the Marsh/Swamps and Pan/Saline wetland classes.Comparisons between the random forests visual and automatic classification showed that the area coverage for the lakes/rivers/ponds are almost equal to each other.For the other wetland classes the results of the area coverages varied.The classification accuracy of the visual classification method was higher than that of the automatic classification method.Although its accuracy is higher,the visual classification method is more appropriate for classifying or determining only the lakes/rivers/dams/ponds wetland class.This is because the accuracy of training samples for visual classification is dependent on the experience and eye-sharpness of the remote sensing specialist establishing the training samples.Other than the lakes/rivers/ponds wetland class,the other wetland classes are not easy to identify on a satellite image.The study also determined that the proposed automatic wetland classification method which is landcover-based is more appropriate when considering a broad range of wetland classes as well as a very large areal extent.The area coverage of wetlands in Africa for the period 2001-2017 either increased or decreased with each year because of the different annual rainfall for the different years.The variation in the distribution of wetlands was found to be influenced by factors such as population,economical activities,precipitation and climate change.In future,to improve the classification accuracy,different landcover datasets can be incorporated.Ground data to validate the results of the classification system can also be collected in advance and different vegetation and water indices can also be used to determine how different wetlands vary according to these indices depending on their plant communities.In addition to Landsat,other satellite imagery such as Sentinel which has a radar band can be used.The advantages of using radar is that it can penetrate under water.In regards to this study,future work should be on improving the accuracy of classifying wetlands and improving the remote sensing-based wetland classification system.Future prospects of this study lead to a global wetland classification maps that leverages on platforms such as Google Earth Engine.