Determining the presence and extent of surface water bodies is of paramount importance in order to monitor mosquito breeding habitats. With the availability of high spatial and temporal resolution SAR images through the Sentinel-1 mission, it becomes possible to map surface water on a regular basis in mosquitos' prone areas. In this work the potential of Machine Learning (ML) algorithms is investigated for the near-real time mapping of surface water in the Chalastra plain in Central Macedonia in Greece, using Sentinel-1 data. Three ML algorithms: One-Class SVM, One-Class Self-Organizing Map, and Multilayer Perceptron with Automatic Relevance Determination (MLP-ARD) were compared to the Otsu Valley-Emphasis method, a commonly used approach based on histogram thresholding. All methods were automated and tested in the pilot area. Results show that the MLP-ARD algorithm achieves the highest overall accuracy (0.974) among all methods with a kappa coefficient of 0.933.
The distribution of surface water bodies varies with time and space depending on weather conditions, irrigation patterns, as well as different water uses. Regular mapping of water bodies is a good practice towards the sustainable management of water resources. Water ponds, wetlands, inundated rice, and flooded areas are some major habitats of mosquitoes’ breeding. Monitoring the presence of water helps identifying hotspots of mosquitoes’ larvae development, for targeted and effective larvicide application, in order to reduce their potential harm to human health. With a rational use of such larvicides the impact on water resources will be minimized.
Remote sensing data, both radar and optical, were widely used for monitoring
surface water from space. Optical satellite data were used to produce surface water
indicators (e.g. the Normalized Difference Vegetation Index - NDVI, the Normalized
Difference Water Index - NDWI and the Modified NDWI - MNDWI)
There is also an emerging trend in the literature
In order to identify pixels with water on satellite images, histogram thresholding
was widely used with multispectral and SAR data
ML combined with remote sensing techniques have been utilized widely in order to detect temporal and spatial alterations in small waterbodies so as to provide information in the field of water security, drought monitoring, and crop irrigation decision-making. The above combination forms a reliable, cost-effective, solution for surface water mapping in large and inaccessible areas. Bangira et al. (2019), proposed a such an approach based on the effective combination of automatic thresholding of NDWI using Sentinel-2 and backscatter from Sentinel-1 data with VH polarisation for mapping waterbodies of diverse spectral and spatial characteristics. The resulting maps were compared to the classification performances of five ML algorithms, namely decision tree (DT), k-nearest neighbor (k-NN), random forests (RF), and two implementations of the support vector machine (SVM), enhancing the accurate classification of optically complex waterbodies.
In the present work three ML algorithms are presented and tested against a classical method of histogram thresholding for the automatic mapping of surface water bodies in the Chalastra plain, Greece.
The study area is the plain of Chalastra, which is the lowest part of the floodplain of rivers Axios and Loudias (Greece). It is intensely cultivated with irrigated annual crops, most predominantly with paddy rice. Numerous drainage canals, river segments, and their estuaries have formed a complex wetland system (Fig. 1).
A field campaign was conducted during summer 2019 to collect ground data of wet and dry polygons in the wider area around Chalastra. The samples were taken at dates concurrent with the overpass of Sentinel-1. Sampling locations consisted in rice fields (inundated during the early season and presenting different plant growth stages) and wetlands which are typical breeding habitats but also built areas, natural vegetation and dry soils which can challenge the water detection algorithms and help training the ML ones. The mapped polygons were characterized by different water depths and vegetation cover. On the date of June 12th 2019, a total of 30 polygons were collected leading to a total of 16866 10-by-10m pixels labelled as wet or dry in the test area.
A Sentinel-1 image was downloaded using Google Earth Engine (GEE) with both VV and VH polarizations, Ground Range Detected (GRD) type and in Interferometric Wide swath (IW) mode. The downloaded image was already pre-processed in GEE as proposed by the Sentinel-1 Toolbox and saved as Geotiff data file with the backscatter given in decibel. A section of the image was used for processing (2552*1102 pixels) (see Fig1).
Four surface water mapping algorithms were implemented in Matlab and tested on the Sentinel-1 image of June 12th, 2019. In particular, three ML algorithms were compared to a revised histogram thresholding method.
To achieve One Class SVM Classification, a suitable description of SVM as a model
to describe only target data introduced by Tax and Duin (2004) in the form of Support
Vector Data Description (SVDD) has been applied. The OC-SVM develops a model
by being trained in using normal data conforming to the SVDD description. At the
second stage, it allocates test data based on the occurring deviation from normal
calibration data as being either normal or outlier
It can be determined by considering that a sizeable spread indicates a linear class of target data while on the other hand, many support vectors joint with a small spread indicate a highly nonlinear case as is illustrated in Figure 2. A spread parameter equivalent to 1 yielded the best results concerning waterbodies detection. The threshold for accepting outliers was set at 10 %. 3.2
The OC-SOM constructs a model from target data (dry areas) and progressively
classifies new data conferring to its deviation from the target baseline condition.
During novelty recognition, novel instances from feature combinations (in the current
case the features from the S1 image) of not definable state validation data samples, are
used to form the input to the network, while the SOM algorithm chooses the Best
Matching Unit (BMU). In the current application, the BMU is defined as the most
proximal SOM centroid vector in the weights space to the incoming data vector. In the
occasion that the quantisation error that is the outcome from the appraisal between the
new exemplar data (xNEW) and BMU is larger than a pre-specified threshold (d) then
the example is considered as novel
MLPs are feed forward artificial neural networks, that in classification problems map a set of input vectors onto their respective classes. For this study, a fully connected MLP with a three – layer architecture (input layer, hidden layer and output layer) is assigned for the classification of the spectral signatures into the healthy or the intensity of the disease level conditions. The weights correction is performed by the scaled conjugate gradient back propagation algorithm and the transfer function that were selected were the hyperbolic tangent (tanh) for the interconnections between the input and the hidden layer and the logistic function for the respective interconnections between the hidden and the output layer.
Apart from the first level hyperparameters, the values of which are randomly chosen
as priors for the initialization of the MLP classifiers, Automatic Relevance
Determination (ARD) is used in this study. In the application of the ARD technique a
new regularization hyperparameter, alpha (α) is introduced, for every weight that is
associated to the i input variables, in order to determine the relevance of the input data
into the model
The Otsu Valley-Emphasis method
For the implementation and assessment of the ML algorithms the field survey dataset was split. 11806 samples were used for training and 5060 for validation, while the Otsu Valley-Emphasis method was validated using the whole dataset. All four methods were implemented using Matlab version R2019b.
The results of the best performing ML algorithm the MLP-ARD are presented in Figure 3.
The MLP-ARD method led to the highest overall accuracy of 0.974 with a kappa coefficient of 0.933. The precision and recall metrics for this method are respectively 0.987 and 0.977, higher than for the other methods meaning that false alarms and missed detections are minimized.
The other two one-class ML methods OC-SVM and OC-SOM provided a lower accuracy than the Otsu Valley-Emphasis method. However, in precision and recall OC-SOM algorithm supersedes OVE. Both OC algorithms have equal or better recall than the OVE. This result can be explained by the OC algorithms ‘excellent outlier detection capability with the recall of OC-SVM reaching a performance of 0.986.
As shown in Fig 3. the sea, riverbed and coastal wetlands were detected by the MLPARD algorithm with few missed detections. Unlike for the one-class algorithms, the built areas in the neighboring villages did not cause false alarms.
The computation time required to train the OC and MLP -ARD algorithms is minimal and requires very few minutes. For producing a prediction map for waterbodies detection with a size of over 2.8 M pixels, different time was required for the two OC algorithms (few hours), which appeared to be a limiting factor for nearreal time application. Both the MLP-ARD and OVE methods produced maps in a reasonable time (few minutes).
Both the Otsu Valley-Emphasis and MLP-ARD algorithms provided good accuracy for the detection of water bodies using high-resolution SAR images. Due to the generalization capability of the MLP -ARD, the trained network can be applied further without any retraining and can generalize over new example images acquired from different days and years. A more robust version of the MLP-ARD will be produced by training by data from different days’ time series of Sentinel-1 images will be processed using the MLP-ARD to monitor surface water bodies during the whole summer season of 2019. Results will be presented in a future communication.
Acknowledgments. This research has been co-financed by the European Regional Development Fund of the European Union and Greek national funds through the Operational Program Competitiveness, Entrepreneurship and Innovation, under the call RESEARCH – CREATE – INNOVATE (project code: T1EDK-05106).