scale.bythebay.io: Alex Kudriashova, Satellite data monitoring for analytical models
Recording: scale.bythebay.io: Alex Kudriashova, Satellite data monitoring for analytical models
about satellite data for monitoring and do it for analytical models for some reason I know nothing about Scala and they stack for our company is little different but still we have the same challenges they main problem for us is how to process this massive amount of data at scale so the satellite imagery is really cool you can persistently monitor the dla any location with different sources of data and see the change so for example the image behind me is not a static picture taken from one satellite shot it's actually encoded time series of images so for every pixel we tracked how the behavior look like through agricultural season and this is corn fields for Iowa and for the areas colors with the same color we can confidently say that this areas were they were harvested the same moment which means that it's the same crop and also it's very likely to be the same farmer who order the machinery to stand for for a certain location so have an insights like this just purely on satellite data without any knowledge about what action is going there like without any reference data allows us to do the monitoring and get decisions and share these decisions with different stakeholders who would eventually be interested in this data but the main problem here that they later said we're currently talking about is five to six petabytes and the challenge here is how to effectively access this data so what our company is doing we are building we are building an API and the platform for easy access to the data because like no one like not so many people need the really global access everyone is interested in at certain locations which is a business act is actually happening so what we are building is the archetype of the e10 the processing pipeline that goes and looks specifically for the area of interested interested for for the users and we process this data and distribute to the users through yeah which can be easily visualized or downloaded or be used in analytical models on the user side so the first as I said the first part of it is a monitoring we pool all the available data from different sources including the our own constellation our company is built in and a public domain data then we then the analytics starts so what's the first cool thing is there measurement we have not only the current observations but also about 20 years of monitoring and we can come to know that how they what pixel behavior look like this year many previous years and expect how they it should look like for the next year so comparing this pattern where the ongoing with the ongoing observations we can track changes and tell where there is some unexpected activity or off pattern activities happening and with this data is possible to manage to require more high-resolution data or take actions so from these three steps basically they as to digital vision built in so first first to the data so just to get a sense of how the satellite images looks usually we are talking about the satellites which fly on the low-low orbit they are following the Sun so solar synchronous orbits so it's like a cloud of different satellites form in the Sun and they take shots so the typical shot is 200 by 200 kilometers and the spatial resolution is 10 to 30 meters in 1 pixel comparing to for example Google Maps it's not a super exciting resolution but still if you have a consistent imagery image daily for the entire land mass it's really a valuable data source maybe not only for visual assessment but also for analytical models when we compare the previous day or the previous year with current here and can measure they change the track the same pattern activity let me give an example of how the typical imagery looks like so Landsat 8 is their NASA satellite launched in 2013 which makes the resolution of 30 meters and it has 9 spectral bands it images bill or the entire land mass every 16 days so it's typically two images per month and plus it has the previous generation of Landsat 7 who has archive of 18 years in addition to that with the same spectral parameters European Space Agency launched their Sentinel 2 satellites and they are archive styles in 2015 and it's 10 times more high resolution because it is 10 meters no pixel and the temporal resolution is every 10 days so we have a shot for every 10 days and combine these two guys they allows monitoring every 4 days which is almost like a two shots per week but effectively it's not enough especially for cloudy regions because the problem with multispectral imagery is that it cannot look through the clouds so what you want to do is to make shots as frequent as possible ideally every day or maybe several times a day to increase the chance of having a cloudless image or for example two views and several data sets in order to cover the pixels which were under clouds with one image with the pixels from another image so overall it would generate us something like this so this is a multispectral image of the vegetation index and by the same color again we see the areas with the same type of crop in the same phonological stage or the same phase of growing so we I compared this dataset where the map of vegetation from USDA and for this specific location they were tomato Tomatoes grow in that year and by comparing the tomatoes with neighbor field or with tomatoes videos here we can say that maybe there is an evidence of some disease there but this is more about the visual assessment because what we want to do is to go and beyond pretty pictures and be able to have the measurable analysis to do some more interesting stuff so for example the Builder type actually so here is Dubai and the red dots are the images which were all able to be newly created buildings and the green dots are the images that were suspects to be to be buildings done by the classification why we used by first of all because they are the building process is really active and secondly because it's really a cloudless region so it was a successful project for us and for the green dots they were proven to be suspects to be buildings and specifically for that area they only work which was supposed to do is only to draw a more precise lines from those buildings but actually the fact that something is being built there was proven by this type of limit so if we look more precisely in how this type of operation was done we can see that there was like a small area like a small window was extracted and taken form a series of images actually there were 50 images used then we pixel wise look at the spectral behavior for every location and they take areas which show the behavior which looks similar to newly newly buildings so in the bottom image bottom image these areas were highlighted by purple and we can confidently say that for all the purple areas we see that they are the building's spiritual appeared through this period of two years so yes we detected these images which were built in two years and such type of analysis can be done globally for the entire land mass which is for our constellation of satellites is 15 million square kilometres this is a really heavy computational problem because it means to process 200 over 200 million images per year and if we look at the typical satellite shot we can see that it is like three spectral bands and we have this amount of pixels to process precision of one scene took 30 seconds so if you take a typical unknown and on my laptop the processing of this just a daily data set with took 17 years but the problem with the real problem with this data is there the human activity and the areas of interest which are interested to humans is not equal to distributed all over their land mass we really there needs a less than 1% of data to be processed but the question is what exactly the location I needed and how to process these locations cheap effectively and really quickly because another problem is that for many applications the data should be delivered really quickly so for example for hedge fund for trading that data is required the same day so like eight hours window including the time to downlink the data so effectively for the processing of data we have hours maybe minutes so what can we do specifically for our case we introduced the window base treat architecture for AWS lambdas so what we did we define a problem and analytical model to the process define the size of the window and the depth of the archive to look at we feel this too required number of AWS numbers and perform their analysis so this let me go a little bit deeper how it looks like so we don't download the process the full scene because it's like a one gigabyte archive or just one data point while we need to observe thousands of them so what we do is that we send a specifically bounding box around the area of interest for their location to be processed and in percentage like one percent of the image or maybe less so we work on the with this small chunk of data and for that there are limitations for AWS implementation which is about one and a half gigabyte of RAM and 500 500 megabytes of memory of storage is enough to do the process and the another cool thing that we can do it in parallel and the most important part is that their satellite archives they are distributed within the different data centers so we can load locate the this lambda functions in the same data center as they as the data is located so having this we don't waste too much time on sending the data from one data storage to another and the like it's really a good idea to have the processing capacity near to the data to optimize the for sending the data so this implementation allowed us to do the to proof in the following directions first of all we can do 10,000 calls per data center which means that 10,000 threads of different tasks for processing so having these it's possible for us for example to assess their to make a crop classification for the entire corn belt for the one run of the of the system which takes but only 30 minutes all together so after 30 minutes we know we can get the classified crops for the entire corn belt right away comparing to current implementation it is like saving not hours a days another one important thing for us is the cost models so what will cool about the function as a service let me pay per run so depending on how many required how many indications we are charged for which makes us to align the actual cost model with the implementation model another one is the quick deployment since we are a venture funded company we need to grow fast and we need to make sure that the implementation and every iteration is deployed really quickly so that potentially we would consider some other more in-house implementation that country their cloud server last version of the infrastructures totally fine for us as well has a good login and traceability so we can track for example what went wrong and yeah as they conclusion to how to oppress so really the settler data is a fed data of about 5 petabytes and it requires really resources the infrastructure and being smart about what exactly is going to be processed so they function as a service there was a good solution for us and they as to digital monitoring platform as online so we have a web browser so you can draw an area of interest and get the series of satellite images to look at so feel free to check out or to talk to me how to use the satellite data for different tasks Inc yeah [Applause] any questions about satellites or AWS yeah so in it in addition to Landsat and sent no answer in addition to the Landsat and sent know our own constellation of satellites it has the same spectral bands to her to have a consistent time series but it provides the daily monitoring so the fleet of 10 satellites monitors their entire land mass once everyday thank you [Applause]