Faecal sludge characteristics are highly variable, with faecal sludge influent to treatment plants being up to two orders of magnitude more variable than municipal wastewater, with the characteristics following a much different distribution. The high variability presents a unique challenge in operating and monitoring treatment technologies, and makes predictions of characteristics for sanitation planning in cities difficult. Characterization of treatment-relevant metrics, such as total solids, ammonia, and dewatering performance, are also currently difficult to achieve in many places that rely on onsite sanitation, because access to analytical laboratories is prohibitively expensive or simply not available. In situations like this, it would be helpful to be able to supplement laboratory analysis with fast, cheap approximations of important sludge characteristics.
In practice, operators and researchers already use their expert knowledge to predict sludge characteristics and adjust operations accordingly. For example, sludge color as a predictor of level of stabilization, or whether the sludge comes from a pit latrine or a septic tank as a predictor of total solids. So we asked, how can we quantify and improve these “field predictors” of faecal sludge characteristics to make them more widely useful and applicable?
That was the focus of our recent Water Research publication, where we tested a number of possible field predictors based on practitioner experiences, such as taking a picture and using the color of the sludge as a predictor, and quantified their power to predict treatment relevant analytical parameters using machine learning models. Results of this study were quite promising. Machine learning models were a substantial improvement compared with current estimates of practitioners. For example, using operator knowledge such as pit latrine or septic tank to predict total solids gave an R2 of 0.2, whereas the random forest models we generated based on photographs (color and texture) and probe data worked much better, with a higher prediction accuracy R2 of 0.6.
To fully scale up, we need to collect data in new cities and use it to build a global data base of sludge characteristics. Based on that data, we can understand what models are globally or locally applicable, and better quantify prediction uncertainties when the models are used in new cities. We also plan to streamline the data processing in the app to make it simpler and faster for users in the field. For more details, please watch the video, and read the following publications.