White Water relies on a deep neural network that takes in inputs from various samples of polluted water in order to identify abiotic or microbial particles. After being trained various types of bacteria and particulates indicative of water pollution, the app accurately names particles as one of the two types. The current model has a 90+-5% accuracy rating – higher than most commercial water quality analyses – and is capable of use in underdeveloped communities that cannot afford tradition laboratory-intensive methods of identification.
In my research, I learned that it was possible to target one of the most devastating means of death in the world – neglected zoonotic diseases. More than 60% of all preventable deaths are caused by one of 7 of the most deadly water-borne illnesses, which include malaria and schistosomiasis. While these diseases are often easily treated, they occur most frequently in areas of the world where victims are unable to afford even the simplest means of survival.
By conducting an analysis of all solutions, I discovered that the primary issue was that it was difficult to find a way to transport laboratory equipment to underdeveloped communities. The process of water quality analysis was often time intensive, preventing water samples from being sent overseas in a timely enough matter to merit the costs and resources it took for the identification to occur.
From my knowledge of computer programming and inspiration from PyCon 2016, I decided to utilize machine learning and pixel classification to help solve this problem.
What inspired me/us?
After attending PyGotham last summer, I was curious as to whether I could use Python to independently conduct a machine learning analysis on data to help solve the water quality analysis crisis. I understood that some of the world’s most fatal waterborne illnesses were perpetuated by an inability to easily recognize particles in water samples, but current efforts in the field were limited to increasing the viability of vaccines or the distribution of cures for such diseases, as opposed to preventative measures. After contacting professors at Stanford and Columbia University, I received information that ongoing efforts did not specifically target image-based recognition, and I was curious as to whether it might be possible to determine the legitimacy of such a method on my own. As a result, I began to use data that was available about the visual parameter of particles – both abiotic and microbial – and applied an image recognition algorithm to examine the differences to see if I could develop a technology that would ultimately be able to differentiate the different types of particles from a photo sample of water.