For many years now, my experience as a GIS and remote sensing specialist has been marked by projects in the humanitarian sector: the UN, World Bank, Red Cross, WRI, Translators without Borders; you name it, I probably did some work for them. Back in 2018, when I worked for Urban Resilience Platform – a French startup dedicated to developing software for solid waste management in humanitarian crises – , we were given a very simple but hard to accomplish task: help UNDP and the Turkish government identify every illegal dumpsite in Turkey.
But why, though?
Turkey’s unique geographic position with a 911 Km border with Syria, and its standing as a land migration route to Europe has resulted in the country receiving a large influx of Syrian refugees. Since the beginning of the Syrian crisis in 2016, millions have fled across the border to the Southeastern Anatolia region in Turkey. In July 2018, the number of registered Syrians in Southeastern Anatolia was over 3,5 million.
This has led to rapid change in the demographic structure in several Turkish border cities. In Kilis, for example, the number of Syrian refugees is greater than that of the local population.
As populations move, so does the waste they generate. With the doubling of domestic solid waste quantities, Southeastern Analotia’s provinces began to face ever-increasing issues with inadequate disposal and insufficient capacity to collect and treat all the waste being generated by its new populations. That led to an imbalance in the solid waste value chain, resulting in the apparition of many new illegal dumpsites to receive the waste that could not be processed by the municipalities.
UNDP, with its clear mandate to support municipal waste management, responded with the Effective Urban Waste Management for Host Communities Programme, which incuded key infrastructure investments, like waste transfer stations and recycling facilities.
But the diversion of waste away from uncontrolled dumpsites is only the first step in the upgrading of the whole waste collection system. UNDP also wanted to ensure those dumpsites were on course for rehabilitation. Hence the need for a methodology to identify dumpsites.
How do you process imagery for a whole country?
Since we didn’t have the means to travel the country in the hopes of bumping into well hidden dumpsites, we resorted to the use of (drum rolls…) satellite imagery!
Folks in the remote sensing field have been working with machine learning techniques way before it became the hot stuff it is now. The main difference is that they’ve been doing it in a nearly analogic way, with a lot of button clicking involved. The process usually looks something like this:
- Decide which satellite imagery provider you’ll use, usually with the help of a system like Maxar’s Discover;
- Acquire each image you’ll be needing, by downloading one by one from the server or buying it (which usually takes two or three e-mail exchanges);
- Train desktop softwares like ERDAS Imagine, eCognition, and ArcMap to recognize whatever land feature you’re interested in.
Well, it’s all fun and games until you have to download and analyse images for something as large as a country.
That would require you to download a lot of images, build a mosaic with them, set the right parameters and count on your machine to be able to process it in a reasonable timeframe.
Thankfully we can now run all of these processes with fairly simple scripts, and better yet, on the cloud. That’s what Google provides with its free-to-use Earth Engine tool.
The right spectral bands for the right objects
If you’ve ever worked with image classification before, you know that one should always have a good understanding of the composition, form and texture of whatever you’re looking at. The components of waste however, have different densities, overall sizes, chemical composition, and therefore spectral signatures. Simply put, that is to say, our targets reflect different colors (or light wave lengths) in different intensities, making it difficult to detect dumpsites as one single kind of object.
So we decided to make use of an Assisted Classification Model. This – very classic – model consists of "teaching" our code what dumpsites look like with training samples in a collection of satellite imagery, and telling it to search for more. That means in order to get started, we must first decide which image collection we’ll be using. Since we don’t have any budget for imagery acquisition, we’ll use the best free imagery we can get: Copernicus Programme’s Sentinel-2, which provides for a decent spatial resolution (20m pixel, 10m if you use pan sharpening) and an even better spectral resolution (13 bands ranging from 443 to 2190 nm).
In order to do that with GEE, we declare a variable that we’ll call SENTINELand then use the ee.ImageCollection method stating we’re looking for the COPERNICUS/S2 collection. We can then filter this collection to find images that actually can be of use, by setting thresholds to the percentage of clouds (<10%) and date of acquisition (Jan-2018 to Aug-2018). We also used GEE’s Assets tab to upload a zipped shapefile containing our area of interest which we called southeasternturkey and then applied it as a bounds filter.
Now this is where it gets interesting. Even though we just filtered the collection, our SENTINEL variable still holds a lot of images, all superposed and sometimes overlapping. It would take us a really long time to pick the best images and build a mosaic with them. But GEE provides the very useful median method, which provides the median pixel value for any given point in space, taking into account the filters we have already provided. So we add that new SENTINELMED mosaic to the map and see what it’s natural color RGB (B4, B3, B2) composition looks like.
Is that waste or just rocks lying around?
Ok, great! That’ll be the mosaic we’ll be using to try to find our dumpsites. What we need to do next is establish some training samples, so our code undersands which pixels are dumpsites, and which aren’t: water, forests, crops, cities, exposed soil, etc. For the dumpsites, we’ll just import the shapefile with some previously known dumpsites provided to us by the UN, and for the rest of the landuse features, we’ll be drawing those directly on GEE’s map. Once a new layer is created in the Map pane, it shows up as a variable in the Import pane. It’s essential that both our imported data and our drawn features have a property, which we’ll call class, and an unique value. For instance, dumpsites are gonna be 'class:1', water will be 'class:2', and so on.
Once we have collected enough samples, ranging both in categories as in extent, we can merge them all into a single sample variable. We also need to establish which spectral bands will be used to perform the classification. In our case, we’re using all of them. Next, we need to tell GEE to use our geometries and extract the pixels from the chosen collection in order to establish sample regions. This is the moment where our class property is used.
Now we can use the Classifier method to train an algorithm. GEE has many different classifiers, like Decision Trees and Random Forest, to cite two of the most populars. Here we’ll be using the classic Cart classifier.We’ll feed it our sample regions and chosen bands, and it will return our own custom classifier, which we can use with any image collection we want.
Finally, we can create a classification variable in which we’ll tell our classifier to do its magic with our SENTINELMED mosaic. We can then add the result of this classification to the map, stating the colors for each class we created.
In order to do some more processing, we then needed to export the resulting raster file to our local machine. GEE makes it easy by providing an export to Google Drive function.
Now what?
The classification of pixels is not error proof, especially considering how different dumpsites can be from one another. For this reason, the model requires a last step which consists in analysing, by visual observation, clusters of reclassified pixels, as shown below. We could have used a segmentation routine – that is, transforming our mosaic into single color geometries – before having run the classification, but we felt that since there weren’t that many dumpsites, the human eye would be more precise in correctly identifying them.
The model is most effective when it comes to identifying dumpsites receiving domestic waste, as the dark colour the organic matter assumes is very characteristic. Inversely, when it comes to dumpsites receiving rubble or waste from construction or demolition sites, the task of identifying the clusters gets harder, as those can easily be mistaken for mineral extraction sites. For those cases, the visual analysis is most important, as the patterns observed in this kind of dumpsite are better recognized by the human eye than by pixel classification.
In total, seventy polygons were delineated through this process. Each of these polygons is suspected to be a dumpsite. Nonetheless, some uncertainty remains for some sites, which should still be further verified. For that, we created a WebApp using an ArcGIS Online app template for data editing. The app, shared with UN and Turkish Government officials, allowed each user to provide input by classifying each delineated polygon as (1) Construction waste, (2) Domestic waste or (3) Not a dumpsite.
With the input from key informants, we can now finally say we successfully identified every dumpsite in Southeast Turkey. This project, conducted in August 2018, was the first step UNDP took toward rehabilitating the illegal dumpsites in the region. What follows is an intense on-the-ground set of actions to ensure that all the stakeholders in the waste value chain, from informal collectors to recyclers, are aware of the new available waste treatment structure. Control actions undertaken by the Turkish government can clear the way to effective dumpsite rehabilitation.
You can find the Google Earth Engine image classification model by following the link below. If you’d like, don’t hesitate to drop me a line here or on Linkedin!
