Microbiological Water Quality Test Results Extraction from Mobile …jeffxing.com/files/iiwas_presentation_water.pdf · 2020. 5. 23. · Motivation Related Work Methodology Performance

Motivation Related Work Methodology Performance Conclusion References

Microbiological Water Quality Test ResultsExtraction from Mobile Photographs

Ngurah Agus Sanjaya ER1, Jifang Xing2, Zhang Ruixi2,Remmy Zen2, Laure Sione3, Ismail Khalil4, Stephane Bressan1

1Udayana University, 2National University of Singapore, 3Imperial College London,4Johannes Kepler University

April 20, 2020


Motivation

The United Nations have identified access to water andsanitation for all as the sixth of seventeen goals in the UnitedNations Sustainable Development Goals.

This is a hard task to achieve in the developing world due tothe effects of climate change, compounded by a lack offinancial resources and trained personnel.

Intermittent Water Systems provide water discontinuously intime and space, which not only results in an unreliableprovision of water, but also one that is unsafe for directconsumption (Waterborne pathogens, contaminatedgroundwater).

An emerging and promising approach to tackling these chal-lenges is citizen science which has been successfully applied innatural world [EC19].


Motivation

This paper specifically considers water quality data collectedusing a mobile phone app wielded by citizen-scientists. Thecitizens use a microbiological water quality test kit to measureE. coli content. The test result is photographed by thecitizens and passed on to scientists for interpretation.

However, reading the results necessitates a trained scientificeye and is time consuming. This paper therefore puts forwardan algorithm that can automatically infer the outcome of thetest from a photograph.

Specifically, we consider the collection of water quality datausing a similar crowdsourcing technique to Networking WaterData project [wat19], namely the Aquagenx [Aqu12]microbiological water quality test (E. coli) kit [SETP19]. Thekit consists of a compartmentalised bag and a growth medium.


Motivation

Figure 1: Photograph of the bag

Figure 2: MPN table.


Goal

Water quality test

most probable number (MPN) of E. coli1 fromAquagenx [Aqu12] water quality test (E. coli) kit [SETP19]

colour combination of five compartments determine MPN

32 colour combinations totally

1a proxy measurement for water quality


Related Work

mWater [mWa12] and Akvo Caddisfly [Akv08] that allowusers to upload photographs and manually enter water qualityresults.

The inference could be based on the detection of thecompartments in the photograph followed by the inference ofthe main colour of each compartment (Canny edgedetector [Can87], Deep Contour [SWW+15], Active contourmodel [KWT88]).

The inference could be a more direct classification orregression tasks that tries and directly infers the test outcomeor the most probable number of E. coli [GBC16].


Classification

Two ideas for classification

rule-based inference:– project yellow and green regions to x-axis and y-axis– design rules to infer colour combination of the compartments based

on the projections

neural network classifier:– input: processed images– output: possibilities of 32 colour combinations (32 output nodes,

CNN-32) or possibilities of yellow outcome for each compartment (5output nodes, CNN-5)

– use classical model of LeNet [LBB+98]


Problem Definition and Framework

Problem Definition

input: photograph of water quality test (E. coli) kit, denotedas X three-dimensional matrix with dimension M ×N × 3,where M and N are the width and the height, respectively,and the three layers corresponds to the red, green and bluecolour of the photograph in pixels.

output: predicted MPN

key point: classification c = F (X), where c = [c1, . . . , c5],ci ∈ {G,Y}


Problem Definition and Framework

Framework

Predicted outcome

Predicted MPNpreprocessing classification mapping

original photograph

processedimage

Figure 3: The framework of our method. MPN stands for the mostprobable number of E. coli.


Preprocessing

Figure 4: Preprocessed image for the photograph of a bag shown inFigure 1.

Normalise the photographs by cropping the region thatcontains the bag and resize the photograph from M ×Npixels into a square image of dimension 512× 512 pixels.

The former is achieved by only keeping the pixels with acolour within the range from (8, 50, 10) to (100,255,242).

Isolated pixels or small groups of pixels are further eliminatedusing the erosion operation with appropriate settings.


Rule-based Inference

Example of projection

Figure 5: A 4× 4 image where each pixel is either yellow, green ortransparent.

Projection:Rg

x = [1, 0, 1, 1], Ryx = [0, 1, 0, 1], Rg

y = [1, 0, 1, 1], Ryy = [0, 1, 0, 1]



Figure 6: Boolean test vectors

Figure 7: Rules Figure 8: Output Sets



We design a set of rules to infer the possible colour combination ofthe test kit based on the projection Rg

x, Rgy, Ry

x, Ryy

Example of Rule-based Inference for above figure:

Rgx and Ry

x have value of one, cannot be [G,G,G,G,G] or [Y,Y,Y,Y,Y]

Rgy and Ry

y indict green and yellow regions both exceed 0 to 2H/3 range ofy-axis, cannot be [G,Y,Y,Y,Y] or [Y,G,G,G,G]

Rgx indict green region is within 0 to W/2 range of x-axis, can be [G,G,Y,Y,Y]

or [Y,G,Y,Y,Y]

there are two consecutive regions of Ryx, cannot be [G,G,Y,Y,Y]

Result is [Y,G,Y,Y,Y]


Metrics

Measurement

Aw = 100%× number of correctly predicted outcome

size of the data set

Ai = 100%× number of correctly predicted compartment i

size of the data set

As =1

5

5∑i=1

Ai

RMSE =

√∑i(predicted mpn of bag i− true mpn of bag i)2

size of data set


Data Set

Data Set

Imbalanced and incomplete data set:

251

2 5 9 5 2 6 3 6 3 5 30

50

100

150

200

250

300

0 2 4 8 12 13 16 24 28 29 30 31category

(a) Histogram of outcomes inthe data set.

274 274 271290 292

26 26 2910 8

0

50

100

150

200

250

300

350

1st compart. 2nd compart. 3rd compart. 4th compart. 5th compart.

yellow green

(b) Histogram of outcomes foreach compartment.

Figure 9: Statistical information of data set


Qualitative Evaluation of the Active Contour Method

(a) The detected contourcovers multiple compartments.

(b) The detected contourpartially covers a compartment.

Figure 10: Results of active contour model for the second compartmentwith fine-tuned hyperparameters α = 0.01 and β = 0.01.


Quantitative Evaluation of the Rule-based Method and the Neural Network Methods

(a) Accuracy versus number ofepochs for CNN-32 model.

(b) Accuracy versus number ofepochs for CNN-5 model.

Figure 11: Accuracy versus number of epochs for different CNN models.


Evaluation

Evaluation of Rule-based and CNN Methods

Compared models:Rule-based, CNN-5, CNN-32

Figure 12: Performance of Each Method for Data Set.

Ai is accuracy of prediction for ith compartment. As is averageaccuracy of five compartments. Aw is accuracy of prediction oncolour combination. RMSE is root mean square error of MPN.


Evaluation

Unseen Data

we remove the images of the category 8, i.e., outcome[Y,G, Y, Y, Y ], from the data set to construct a new data set.

Figure 13: Predicted Outcome of Each Method for Unseen Cases.


Summary

We have presented several algorithms for the automaticextraction of results from photographs of microbiological waterquality tests, in particular, we devised a rule-based approach.

traditional image processing algorithms performing on regionextraction, such as the Snake, are not sufficiently robust forthe task at hand.

Neural networks suffer from small and imbalanced data setsand their architecture should be designed carefully.


Future Work

This work is a modest building block in the development ofpractical tools to facilitate the interpretation of water quality data


Acknowledgment

Research reported in this publication was partially funded bySingapore Institute for Data Science under its project WATCHAand jointly supported by the ASEAN-European AcademicUniversity Network (ASEA-UNINET), the Austrian FederalMinistry of Education, Science and Research and the AustrianAgency for International Cooperation in Education and Research(OeAD-GmbH) project ASEA 2019 / Uni Linz / 2.


References I

Akvo, Akvo caddisfly,https://akvo.org/flow-caddisfly-lumen/, 2008.

Aquagenx, Water quality test ki,https://www.aquagenx.com/, 2012.

John Canny, A computational approach to edge detection,Elsevier, 1987.

Economic and Social Council, Special edition: progresstowards the sustainable development goals, United Nations,2019, Available at https://undocs.org/E/2019/68.

Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deeplearning, MIT press, 2016.

https://akvo.org/flow-caddisfly-lumen/

https://www.aquagenx.com/

https://undocs.org/E/2019/68


References II

Michael Kass, Andrew Witkin, and Demetri Terzopoulos,Snakes: Active contour models, International journal ofcomputer vision 1 (1988), no. 4, 321–331.

Yann LeCun, Leon Bottou, Yoshua Bengio, Patrick Haffner,et al., Gradient-based learning applied to documentrecognition, Proceedings of the IEEE 86 (1998), no. 11,2278–2324.

mWater, mwater surveyor mobile app,https://www.mwater.co/surveyor.html, 2012.

https://www.mwater.co/surveyor.html


References III

Amadu Salifu, Helen MK Essandoh, Afsatou Ndama Traore,and Natasha Potgieter, Water source quality in ahenemakokoben, ghana, Journal of Water, Sanitation and Hygiene forDevelopment 9 (2019), no. 3, 450–459.

Wei Shen, Xinggang Wang, Yan Wang, Xiang Bai, andZhijiang Zhang, Deepcontour: A deep convolutional featurelearned by positive-sharing loss for contour detection,Proceedings of the IEEE conference on computer vision andpattern recognition, IEEE, 2015, pp. 3982–3991.

Networking water, Networking water data project, 2019,Available at https://www.networkingwater.com.

https://www.networkingwater.com

Microbiological Water Quality Test Results Extraction from Mobile …jeffxing.com/files/iiwas_presentation_water.pdf · 2020. 5. 23. · Motivation Related Work Methodology Performance

Documents