 Okay. Very good afternoon to all the people in Europe. My name is Kai, actually you can just call me Kai. Those people in Netherlands, they call me Kai. So I work for the Singapore National Water Agency, CPUB. And I am also a very proud student as what I would have mentioned of IHE, Hydroinformatics Class of 2017-2019. So I just want to share with you very quickly one of the projects that we have done, one of the pilot projects that we have done actually. And I originally had about 44 slides but I have cut it down to just 10 of them. So I just want to give you a gist of it. If you really want to know more details, then I think you can contact me or you know, Abraham for more details. Okay. So what we actually did is we applied deep learning models in flood forecasting for a case study of Singapore. So the deep learning models here actually adopt something like an image recognition and pattern recognition kind of a model which is quite interesting because it's not very common for flood forecasting models. I'll move on. I'll give you a very quick introduction of Singapore. It is a very, very small and tiny island in the Southeast Asia. And a lot of people say that you know Singapore is a city of China. It is actually not. It is a country of its own. And yeah, we are in the Southeast Asia 716 kilometers square which is super, super small compared to other countries. And Singapore just like many countries in the world, we grapple with the challenge of changing weather patterns. Over the years, we have been experiencing higher frequency of rainfall and also of course the magnitude of rainfall. And apart from that, there is also high degree of urbanization. So we are constantly competing for land spaces to house our drainage infrastructure. Singapore is actually very small, so we need to put in all the you know, infrastructures, houses and all that. And there's always a constant need to compete with all the drainage infrastructures. So we recognize that actually flood forecasting is very important. We need to start moving towards you know, very adaptive, non-structural measures as we move on to the future. So there is a flood forecasting system in Singapore just like many countries in the world. And for Singapore, we actually comprise of four systems, four components. The first component is an S-band weather radar. What it actually does is it scans the wind clouds. And then with the radar images captured, it is then being fed into a now-cast model. I'm not sure whether you can see something wrong with the presentation right now. So with the radar images captured, it is being fed into a now-cast model. So what this now-cast model will do is it will take in all your radar images and then it will extrapolate them into future radar images. And then it will then translate them into future rainfall predictions. With the future rainfall predictions, you then fit it into a physically based model. And this physically based model is something like your soil bag or your mine level model. So it actually gives you water level at various locations of the catchment. And then we also have a small, you know, error correction module that will help us correct the error. So you'll have water level forecast based on this setup. So what is the problem? So the problem for us is actually at the rainfall now-cast model. So Singapore is actually in the tropics. This rainfall now-cast model that we have adopts something called like the optical flow constraint principles for extrapolation. So it actually assumes that storm cells remain constant in intensity and their movements are sequential. But we know that of course, you know, especially people in the tropics will know that in reality, this is not true because storm cells can grow and you can decay within minutes. Sorry, I think we need to be a little bit faster. So this has actually led to low accuracy in the rainfall now-cast. And poor accuracy in the water level forecast is as well. So this is one problem that, you know, Singapore has faced for many, many years. So over the years, there has been advancement in artificial intelligence. I think Jero shared that with you as well. So mainly in two areas. The first area is image recognition. I think you can sense it as well that, you know, your image recognition kind of tools are getting more and more prominent and more and more powerful. So this image recognition technique that usually people use, they use this algorithm called the convolutional neural network. So what this algorithm does is it's actually very good in capturing and picking up important features. Okay. And the other area that has significant advancement is the recurrent neural network. So recurrent neural network, you know, the algorithm, the backbone behind all those, for example, like your Google Translate speech recognition and all that. So it's actually very good in predicting sequential kind of stuff. So there are significant advancement in these two areas. And you ask me, so what, what has it got to do with, you know, Singapore, right? So in 2015, a group of Hong Kong Met Office researchers, they tried to combine the convolutional neural network with the recurrent neural network. Okay. So it actually adopts the best of both image recognition features. And then we've spatial and temporal kind of prediction for the recurrent neural network. So it combines these two algorithms. This hybrid algorithm is called the ConvLSTM and they apply this ConvLSTM on rainfall now casting problems. So what they did was they used like past radar images and then they predict future radar images of the next 19 minutes. And this Hong Kong Met Office, coincidentally, they also use the same optical flow constraint principle, same as Singapore, for their now cast model. So they are able to compare this deep learning model with the optical flow constraint model that they have. And they found that this deep learning model ConvLSTM actually outperforms the optical flow constraint now cast model. You can see the results here are quite apparent. Okay, move on. So there are actually many parallels which Singapore can draw from this study. And we wanted to test on this ConvLSTM algorithm for us. So what we actually wanted to do is we want to construct a deep learning model which takes in radar images of past 30 minutes. And then we want to predict water levels in the canals. Okay, this is essentially the whole objective of it. So if I were to move on to the previous slides that I have, what we intend to do is we actually intend to bypass all this. And we essentially want to, I can't point it, doesn't work. So what we want to do is we want to bypass, we want to just take in the radar images and then predict water level directly. So from step two all the way to the final step, bypassing all the intermediate predictions and the physically based model that we have. This is what we want to do. I'll move on very quickly. So we actually tested on one of the catchmen in Singapore, very small, 20 kilometers square. And it comprises of mainly residential areas and commercial premises and parks. This catchman is called the Burdo catchman. And it's at the eastern part of Singapore. So what is interesting is we need to process radar images. This is not very common in hydrology and hydraulics. So radar images we got it from our Matt office, this S-Band weather radar. And it is captured at every five minutes. So you can imagine it's very big file that we are handling. And the radar images are of one pixel, one pixel of it is one kilometer by one kilometer in resolution. And the radar images are in RGB format. The availability of the data is from 2011 to 2015, very, very big files that we have. And what we wanted to do is to keep the rainfall intensities and the pixels. We want to remove all other background colors. This is not important to us. And we wanted to convert it to grayscale image. This is to reduce the memory space. So this is what we do. Pick up what is relevant to us. And then from there, we convert it to the grayscale image, something like that. And then we trim it. And in terms of other data sets that we are working with, water level sensors, and you can see that there are five sensors in the catchment. Catchment one, two, and four heavily affected by tidal. Sensors five and three are not really affected by tides. And also we use predicted tidal data as well. And we need to tighten coded to synchronize it to five minutes interval. Very quickly, just maybe two more slides before we end. We constructed many models, but I just want to show you two of it. The first model is we use past radar images, fit it into the deep learning model that we trained and to predict the future 60 minutes water level. So what we can see, give you a good glimpse of it, is that there is actually good prediction for sensor number three and five. But we found that the Rumi square error and the R square that we have the predictions for sensor one, two, and four is actually quite long seed. So you can see the comparison right here. Sensor five, good results, but sensor four, not that good. And what we found is that actually tidal data plays an important role as well. So we flattened the tidal data together with the radar images to predict the future 60 minutes water level. And once we have done that, we can see that actually in terms of prediction results, it does give you the sinusoidal tidal pattern as well as giving you good results for that. So very quickly conclude. The conclusion here is that actually using past radar images, the deep learning models can actually forecast water levels and achieve satisfactory results. And here, actually, the challenge that I have was more on data processing and manipulation is very, very challenging. And but it is also one of the most important component that we should always emphasize on. And lastly, from the tidal information that I shared with you, understanding the area is actually very important. Sometimes there are some parameters that you miss out that would significantly affect your results. Yeah, so that is all I have actually questions, anybody, or maybe we should.