Dataset of Stagnant Water and Wet Surface Label Images for Detection

Clean water is one of the essential things in life. The running water in natural forms is considered as clean water. To avoid exposure to countless diseases, it is imperative to separate stagnant water from clean water. Thus the main objective of the proposed paper is to create an image dataset of stagnant water and wet surface to detect stagnant water. Accordingly, we considered stagnant water images in different forms and sizes to construct the dataset. In addition to that, brown and black earth surface is considered for the wet surface detection. The dataset consists of 1976 labeled images captured from various angles with annotated files. The dataset images are labelled for two classes, namely water and wet surface. This dataset is highly useful for deep learning experts working in the field of disease control management and post-rainfall earth surface monitoring.


Specifications
Computer Vision and Pattern Recognition Specific subject area Stagnant Water and Wet Surface Detection Type of data Image How datapoints were acquired Stagnant water images in different forms such as black, muddy, shiny, and brown were considered for creating the dataset. Images were captured using Samsung Galaxy Note 9 camera with the specifications as below, 12 Mpx, f/1. 5

Value of the Data
• This dataset is vital as it contributes to the future applications of the Sustainable Development Goal -6 of the United Nations, i.e., "water and sanitation" (UN SDG -6) [3] . To achieve accuracy in stagnant water detection, we concentrated on the forms of stagnant water rather than the container type, which makes this dataset unique. • These datapoints are available in the public repository for all the Scientific communities, Research Institutes, Disease control practitioners, and Policymakers. • According to the best of our knowledge, this is the first dataset that provides annotated stagnant water and wet surface images; that could be used in a wide variety of future research studies related to water and wet surface detection. • The data may be reused for conducting experiments related to waterlogging area detection, surface water detection, water sanitation, and potential mosquito breeding site detection. Moreover, the researchers who are involved in earth surface detection may benefit indirectly [2] .

Data Description
The dataset folder comprises separate folders for Indoor, Outdoor and Raw data. The Raw data folder images are unprocessed; hence resolution range is given for them ( Table 3 , Point.5). The images in the other two folders are resized and labelled. The image file name with suffix "s" denotes indoor images. The class file stored in the dataset provides the class name used for labelling. The image count is provided according to the direction of image coverage, location and class for the image samples of the dataset ( Table 4 ).
The dataset contains 1976 annotated images. Each image was stored in JPEG format and accompanied by the txt file's annotation, wherein the annotation labels considered were water and wet surface. In the water image, different forms of water can be seen like muddy, black, shiny, and with reflection. All the images were taken from the top view and the side view. These images were captured during 2019 from two cities, namely Nashik and Mumbai, of Maharashtra, India. Two people were involved in image collection, and each one collected a thousand images, whereas labeling is done manually by one person to avoid annotation mistakes. Table 1 shows resultant annotations for the labels shown in Fig. 2 . The stagnant water and wet surface images with their annotations in the text file are stored in the dataset folder. Based on the image location, the dataset directory structure is divided into two main folders: Indoor and Outdoor. The outdoor folder is divided into three subfolders: a) Water, b) Wet surface, and c) Water and wet surface. Further, these subfolders are divided into two more subfolders: Top view and Side View. The indoor folder consists of Top view folder with water images only ( Fig. 3 ).
According to photo interpretation, samples are annotated using labelImg software, with a label name associated with their class ( Fig. 2 ). For each image, two classes are used with the names wet surface and water class ( Fig. 2 ). For the wet surface class, brown and black earth surface images were considered as they show resemblance with stagnant water form. While for the water class, stagnant water images in various forms were considered ( Fig. 1 ). To cover the entire water body using rectangle annotations, some labels of the same class may overlap. However, rectangle annotations do not overlap if the class is different [4,5] .
The YOLO format stores the annotations in txt file in the following format,

Class X-min Y-min width height
The values 0 and 1 of the txt file ( Table 1 ) corresponds to the water and wet surface class, respectively ( Fig. 2 ). The following next two values are for the x and y coordinates of the annotation, and the rest two are for the height and width of the annotation.

Experimental Design
The image data acquisition process is shown in Fig. 4 . The water and wet surface images were acquired using Samsung Galaxy Note 9 mobile's high-resolution rear camera. In all 1976, images were captured using a camera and then were segregated and saved in respective folders as per their class and location of capture ( Fig. 3 ). As per [7] , we have considered all scenarios while taking images, i.e. indoor and outdoor.
Availability of stagnant water is more during the rainy season, i.e., June and July. Therefore these two months were chosen for capturing images ( Table 2 ).
As water reflects the colour of the surroundings, it becomes difficult to detect water surfaces easily in all forms. Hence this dataset was prepared by considering all adverse conditions required for water detection. To cope with adverse conditions, it is necessary to collect images in different forms such as shiny, black, brown, and with reflection. Hence the images were captured in the morning, afternoon, evening, and late evening. Furthermore, to avoid misclassification, wet surface images are also included in this dataset ( Table 4 ).

Materials or Specification of Image Acquisition System
The imaging system consists of a 12 Mpx Samsung Galaxy Note 9 (F2.4, dual pixel) RGB camera, used with a short exposure time within the range of 200-300 μs ( Table 3 ) [6] . A 15 V battery was used to power all the components of the imaging system. Due to the complexities of terrain, remote sensing data, and classification methods; it is difficult to accurately identify   stagnant water [1] . Thus to overcome the weather and time-dependent variations of illumination in an outdoor environment, the dataset consists of stagnant water images in the form of reflections, shiny, muddy, and black water. While capturing images for wet surfaces, images that were similar to stagnant water were considered.   Table 4 describes the classes, number of images taken and the environments in which images are taken. Images were captured at a man's height using a handheld mobile camera for the outdoor scene, while indoor five images of a dish were taken from the top view by bending down. The images were captured for water flakes and puddles available indoors and outdoors. The photos of the wet surface were taken by considering the water retention areas available around the water body and, in some cases, separate wet surface images collected from other locations.

Ethics Statement
This data is available in the public domain, and no funding is received for the present effort. There is no conflict of interest.