There are more than 60,000 large reservoirs on the globe, more than 10,000 used for hydropower. Thousands more are to come. Hydropower is a major pillar for sustainable electrical energy generation. The way it is manage, however, is everything but sustainable: Reservoirs and hydropower are a threat to millions, if no measures are taken. Earth observation satellite imagery and neural network-powered coastline detection help to unveil this dark side of hydropower.
It is astounding how something as tiny as grains of sand and silt can change the big picture of the planet and the very face of the Earth, telling a devastating story about humanity’s management of supposedly sustainable sources of energy. Unfortunately, most of humanity did not have access to such a big picture for most of its time and would thus not know about those stories. This has changed, though, through the launch of various Earth observation satellites four decades ago and the rise of a neural network driven by artificial intelligence.
With both tools, it was possible for I, along with my team from the renown German journal Süddeutsche Zeitung, to unveil the global and destructive nature of how hydropower and reservoirs are managed to the disadvantage for millions of people around the globe and, especially, future generations. This is how we did it:
As a nonprofit journalism organization, we depend on your support to fund more than 170 reporting projects every year on critical global and local issues. Donate any amount today to become a Pulitzer Center Champion and receive exclusive benefits!
The investigations’ scope: Reservoir Sedimentation and Coastal erosion
The key is understanding the big picture of hydropower and reservoir mismanagement: It is not just water that gets blocked by the barriers that dams pose, but also everything it transports, which are rocks, gravel, sand, silt and a lot of similar sized materials. They would eventually fill up the reservoir, if no measures are taken against it. Many engineers argue that the sedimentation process is negligible due to its slow buildup. However, the science community argues that data, if existent, point toward the other direction: Sedimentation occurs much faster than once assumed.
But data rarely exist for the more than 60,000 large dams that were built on the Earth. Classically, such data about sediment development in reservoirs is gathered by bathymetry, a kind of sonar technique, which uses the reflection of sonar waves. This is, however, rarely done by reservoir managers due to restrictions in budget, knowledge, or manpower.
The consequence of the sedimentation problem is that eventually reservoirs cannot be used anymore for their purposes, be it flood control, potable water or electricity supply, as the sediment blocks all the storage capacity. But damage is not restricted to reservoirs; coastal sediment replenishment also collapses. Thus, the coast shrinks, land gets flooded, people lose their homes and food sources. This is an indirect consequence of bad reservoir management.
The Landsat earth observation program from Nasa was first launched in 1972, the ESA’s equivalent, the Sentinel program, commenced in 2014. They deliver scenic shots from planet Earth throughout various decades. This long period of time makes it possible to compare coastal changes that were caused by reservoirs. The challenge here is to discriminate land from water, which is by no means an easy task. Discrimination of water from land, which can look quite similar, is a task that is prone to errors.
The first step to overcome this challenge was created by the rise of neural network image discrimination tools, which is based on the repeated training of chained mathematical processes called Artificial Neural Networks (ANN) on thousands of prelabeled images, so called Brute Force Learning. This technology has become more sophisticated and easy to use during the last decade, and pre-built, free and customizable ANN networks exist on the web. The work for Süddeutsche Zeitung is based on discriminators from the open source library Scikit-Learn, the MLPClassifier to be precise.
The Multi Layer Perceptron, a variety of the ANN, can be used to categorize, i.e. classify data points, which means it can be used to tell them apart, thus discriminate them. The goal for the satellite imagery was to tell apart land from water and therefore to identify its border, hence the coastline.
The MLPClassifier from Scikit-Learn offers a variety of parameters that build up and tweak an ANN for an optimized classification result on a given set of pre-labeled data. In the case of Landsat imagery, the following build was used:
MLPClassifier(activation='relu', alpha=0.0001, batch_size='auto',beta_1=0.9,
beta_2=0.999, early_stopping=False, epsilon=1e-08,
learning_rate_init=0.001, max_fun=None, max_iter=200,
momentum=0.9, n_iter_no_change=10, nesterovs_momentum=True,
power_t=0.5, random_state=None, shuffle=True, solver='adam',
tol=0.0001, validation_fraction=0.1, verbose=False,
Most of the parameters influence how fast and to which degree operators within the neural network are updated, i.e. trained. The presented settings do pretty much equal the default values at Scikit-Learn and work quite decently on the satellite imagery associated values.
In order to discriminate between water and land and thus identify the coastline, the MLPClassifier needs to lean on a variety of data to recognize the above mentioned variety of possible water and land colors and label correctly water areas. This is done by filtering the satellite images from Landsat and Sentinel for the right (multi)spectral variables, which are e.g. the Red-Green-Blue (RGB)-values of each pixel, short wave and near-infrared values, water and vegetation indices that are calculated from multispectral values and variances. Most of the values are provided per default with the download of the satellite images, which can be carried out using Google Earth Engine’s Python API.
A pre-defined training set that can be used to train the MLPClassifier to discriminate water from land areas can be downloaded here. After the MLPClassifier is trained, it can attribute pixels of new satellite images with categories such as "water" and "sand" with high confidence despite their similar appearance.
Satellite images that are suited for such labeling originate from the Landsat missions 5, 7, 8, and 9, as well as Sentinel mission 2, which means that satellite images are available from as early as 1984 for analysis.
A major drawback of said images is their coarse resolution. For example, Landsat 5 mission has a resolution of 30m x 30m per pixel, which means, that it is not ideally suited for precise coastal change measurements. This is solved by the usage of the Modified Normalized Difference Water Index based on short wave infrared and Green values of the imagery within a certain area of the image, which includes normally a one or low two-digit number of pixels, between which a water and land max value is defined. The usage of the so called Otsu’s Algorithm allows for a precise sub-pixel allocation of the coastal border.
A super straightforward and free application methodology, usable also for similar cases: CoastSat
This all sounds like quite a complicated process and it indeed is. The bad news is that this does not explain all the necessary details. The good news is that one does not necessarily need to know much more, if one wants to apply the described features for a journalistic project that includes the discrimination of land areas and the drawing of borders on satellite images over time, as there are prescripted Jupyter notebooks that can be accessed, modified, and used.
Most of the upper described methodology has been researched and developed by Kilian Vos and his (former) colleagues from the University of New South Wales, Australia. They provided a toolkit that can be accessed on GitHub with all necessary info on how to install and use the tools by creating an environment with Anaconda and subsequently using Jupyter Notebooks.
Vos named the application CoastSat and also provides a suitable set of already labeled training images. It serves to download the Landsat and Sentinel images for the region and time frame of interest. By understanding the basic principles of the section above, it is possible to modify the MLPClassifier and sub-pixel resolution algorithm to adapt it to other projects, where other stretches of land shall be discriminated over time, e.g. rainforest vegetation from agricultural land.
GIS outputs and project integration
For our project, the land/water discrimination was more than sufficient. We identified three story regions (Rio São Francisco, Brazil; Ebro, Spain; Mekong, Southeast Asia) and various illustrational regions for demonstrating the global dimension of the reservoir and hydropower sedimentation problem (e.g. Mississippi, USA; Krishna, India; Kizilirmak, Türkiye) based on a variety of peer-reviewed scientific studies. The sub-pixel bordering could be used for deriving coastlines from 1984 until 2023. Small code modifications were applied to optimize e.g. the size of the region of interest, select suitable satellite sources (Landsat 7 images are often broken), or reduce the number of images downloaded for certain time intervals. The results were converted into GeoJSON files, which are suitable for subsequent usage in geoinformation system tools like QGIS.
Subsequently, manual cleaning and selection of the coastlines from representative years were carried out and the area difference calculated using QGIS. The data was ultimately fused with freely available georeferenced data for river basin extent and hydropower reservoir locations and projected onto the globe within the internal content management system of Süddeutsche, which allows for an interactive approach of interested readers. The reader can freely select between various basins/deltas on a rotating sphere.
The information presented shows how coastlines retreated after the completion of various dams, reservoirs and hydropower facilities around the globe. This retreat is in large stretches caused by sediments, which are blocked by said reservoirs. The damages are therefore not just done on the coastline, but also in the reservoirs. The scale of coastal retreat gives a good proxy of how enormous the sediment trapping within the reservoirs must be.
The sediments, however, do not need to end within the reservoirs. There are methodologies and practices that allow them to continue their flow and to not be blocked by dams. However, those methodologies are very rarely applied, which is a clear and purposeful mismanagement.
To know more about the backgrounds, please have a look at our stories: