by Yudhanjaya Wijeratne (@yudhanjaya)
This repository contains the results of Team Watchdog's work with satellite imagery: 32 satellite imagery maps of Sri Lanka from 2017-2024, ranging from normal RGB, to high contrast versions, to gridded land cover / land use mapped by our in-house machine learning models. The satellite imagery was obtained at a 10^2 meter resolution from Sentinel-2 and is here made available in a commonly viewable format (JPG) at reduced resolution.
What we are specifically interested here is in understanding how Sri Lanka's environment has changed over the years 2017 to 2024, specifically with regard to loss of greenery and the growth of built-up areas and urban spaces. This operates broadly in the space known as land use and land cover classification. While the government of Sri Lanka does do and use land cover mapping, the closest publicly available map is from 2018 as of the time of writing and therefore it is quite difficult to observe a range of years up to today.
In order to be practically useful and usable in a journalistic setting, LULC work must be taken out of a pure machine learning and data science context, and be rendered as images that anyone can open. We've uploaded said imagery here:
These cloud storage drives contain:
-
Satellite imagery from Sentinel 2 for each year between 2017 and 2024. The data in question uses multiple months, specifically three seasons, for all except for 2024, in order to provide a cloudless snapshot of the country.
-
High contrast re-renders of the same satellite imagery. This makes terrain features features far more visible than the ordinary render.
-
A single colour version of the same image, where each pixel of the high-contrast re-render has been processed and its luminosity assigned a value between white and deep blue. We refer to this as bluescale. What it essentially shows is greenery and lack thereof with the darker colours being more untouched greenery and the lighter colours being a loss of greenery; it has the effect of making roads, human settlements, cleared land and other features stand out in brilliant white.
-
For easier reference, land use, land classification map for each year derived using a pix2pix network (a conditional generative adversarial network) trained on the Sri Lankan Land Use Policy Planning Department's 2018 land use map. While not particularly precise, it is much smaller in file size, easier to refer to due to its gridded nature and forms a useful artifact in examining differences across the years.
Due to the very large nature of file sizes involved in this, the imagery available on the cloud storage drives are JPGs. The PNG editions of the same files (double the resolution) total 27 gigabytes; the TIFF versions, which are the original format of the satellite imagery, are even larger.
This repository also contains the notebook that was used for 4). Other scripts used for this exercise have been standardized and uploaded into the team-watchdog/geodog repo, since they are useful units of functionality for future work.
Data used for this project comes from the Sentinel-2 satellites from the European Space Agency. These satellites operate at a 10 square meter resolution for their RGB bands; that is, a single pixel is 10 meters in reality. We used the RGB bands; they are the most intuitive, and we needed to keep processing and data our heads low (early experiments with NDVI, SWIR and near-infrared bands led us to gargantuan data sizes for each single year and comparatively increased processing times).
While originally we obtained imagery directly from Sentinel Hub using their Python libraries, we found it more convenient to use Sepal.io, which in turn uses Google Earth Engine and Google Cloud to set up repeatable workflows that standardize the question of imagery and help us work around disruptions such as electricity or internet outages.
To minimize cloud cover and artefacts, we use data from three seasons of each year up to the 31st of December. The exception is 2024, where, as of the time of writing, it is not yet December, and therefore, we used two seasons of data up to the 29th of July. Imagery used for this pull had to satisfy the criteria of having less than 20% cloud cover; then these are put together into an optical mosaic that gives us a composite image for the year. The general settings for the data pull are here (also available as LK_Optical_S2.json in this repo) and can be replicated on Sepal*.
- note that Sepal requires some setting up and interlinking of Google Earth Engine and Google Cloud.
Imagery obtained this way is tiled into multiple tiffs that are connected together by a single .VRT file. We used QGIS to stitch everything together into a single TIF file for each year.
First, we converted the image to a PNG. At these sizes, conversion using conventional graphical editors gets challenging, so we wrote a small utility that processes an image in tiles: tiff_to_png.py (see geodog repository).
To generate the high contrast imagery, we converted the image into a three-channel (RGB) image, arranged it into two layers, and applied the multiply operation to combine the layers - ie: multiplied the pixel values of the upper layer with those of the layer below it and then divided the result by 255.
To generate the bluescale imagery, we wrote a utility that uses the same tile-based approach as above to calculate the luminance values of each pixel within a tile and map them onto an axis ranging from dark blue ( 0, 0, 50 ) to white. This is blueconverter.py (see geodog repository). This utility also allows us to rescale and adjust contrast as required. We used a contrast setting of 125% and a scale setting of 50% for the final results.
Finding training data for this kind of work is difficult. Usually, in land cover and land classification work, there exists ready-made shapefiles that are the results of surveys done by governments; these form the basis for training and machine learning model. In our case, we had a single image to go on:
Attempts to convert this to machine processable formats led to many of the scripts within the geodog repository. However, this kind of challenge is also an interesting opportunity. You may have seen applications that will re-render an image into the style of another, for example photography rendered in the style of Van Gogh. The approach we took is similar, that is we have satellite imagery, we also have the imagery of a map representing said satellite imagery: could we transfer the style of one to another?
With a great deal of manual processing, we aligned the image to a shapefile of national boundaries and cleaned up the imagery.
- Rocky areas and the names of cities have to be the same color. This would mean that a competent machine learning model would assume that every city in Sri Lanka has an ex to it a very suspicious looking rock formation that spells out the name of the city. These we could not remove (as white also represents wetlands), but we turned them a vivid shade of purple.
- We also removed the red lines that are drawn through the map to represent district boundaries. Although that color is used to signify rubber in the legend, we could not actually find any evidence of rubber, but it was a dead match for the district boundaries.
- We removed the gradient background, the legends, the text around it.
We then tiled this imagery (using boxcutter.py: see geodog repository) into 3614 individual tiles at a size of 256x256 pixels per tile. We cleaned this up by removing all tiles that had our aforementioned vivid shade of purple in them, leaving us with 3400 input-output pairs for training data. The input was a tile of satellite imagery, the output the equivalent tile of the government map.
Using tensorflow's pix2pix example as the base, we built our model training and inference notebook (pix2pix.ipynb in this repository). In a nutshell, this notebook:
- Performs a number of preprocessing steps, including using OpenCV to align each pair of input and output images by matching features between them
- Builds a test and train data set.
- Trains a Pix2Pix model from scratch
We trained this model for 1,150 K steps - a little over 338 epochs. These types of GAN models are difficult to convention analyze through loss values (since generator and discriminator loss values can perfectly balance but still generate nonsense); instead, we observed every 5,000 steps until the model collapse happened around 200K steps later, and tested a variety of different checkpoints until we settled on checkpoint 238.
┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┓
┃ Layer (type) ┃ Output Shape ┃ Param # ┃ Connected to ┃
┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━┩
│ input_layer_2 │ (None, 256, 256, │ 0 │ - │
│ (InputLayer) │ 3) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_2 │ (None, 128, 128, │ 3,072 │ input_layer_2[0]… │
│ (Sequential) │ 64) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_3 │ (None, 64, 64, │ 131,584 │ sequential_2[0][… │
│ (Sequential) │ 128) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_4 │ (None, 32, 32, │ 525,312 │ sequential_3[0][… │
│ (Sequential) │ 256) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_5 │ (None, 16, 16, │ 2,099,200 │ sequential_4[0][… │
│ (Sequential) │ 512) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_6 │ (None, 8, 8, 512) │ 4,196,352 │ sequential_5[0][… │
│ (Sequential) │ │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_7 │ (None, 4, 4, 512) │ 4,196,352 │ sequential_6[0][… │
│ (Sequential) │ │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_8 │ (None, 2, 2, 512) │ 4,196,352 │ sequential_7[0][… │
│ (Sequential) │ │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_9 │ (None, 1, 1, 512) │ 4,196,352 │ sequential_8[0][… │
│ (Sequential) │ │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_10 │ (None, 2, 2, 512) │ 4,196,352 │ sequential_9[0][… │
│ (Sequential) │ │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ concatenate │ (None, 2, 2, │ 0 │ sequential_10[0]… │
│ (Concatenate) │ 1024) │ │ sequential_8[0][… │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_11 │ (None, 4, 4, 512) │ 8,390,656 │ concatenate[0][0] │
│ (Sequential) │ │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ concatenate_1 │ (None, 4, 4, │ 0 │ sequential_11[0]… │
│ (Concatenate) │ 1024) │ │ sequential_7[0][… │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_12 │ (None, 8, 8, 512) │ 8,390,656 │ concatenate_1[0]… │
│ (Sequential) │ │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ concatenate_2 │ (None, 8, 8, │ 0 │ sequential_12[0]… │
│ (Concatenate) │ 1024) │ │ sequential_6[0][… │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_13 │ (None, 16, 16, │ 8,390,656 │ concatenate_2[0]… │
│ (Sequential) │ 512) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ concatenate_3 │ (None, 16, 16, │ 0 │ sequential_13[0]… │
│ (Concatenate) │ 1024) │ │ sequential_5[0][… │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_14 │ (None, 32, 32, │ 4,195,328 │ concatenate_3[0]… │
│ (Sequential) │ 256) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ concatenate_4 │ (None, 32, 32, │ 0 │ sequential_14[0]… │
│ (Concatenate) │ 512) │ │ sequential_4[0][… │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_15 │ (None, 64, 64, │ 1,049,088 │ concatenate_4[0]… │
│ (Sequential) │ 128) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ concatenate_5 │ (None, 64, 64, │ 0 │ sequential_15[0]… │
│ (Concatenate) │ 256) │ │ sequential_3[0][… │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ sequential_16 │ (None, 128, 128, │ 262,400 │ concatenate_5[0]… │
│ (Sequential) │ 64) │ │ │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ concatenate_6 │ (None, 128, 128, │ 0 │ sequential_16[0]… │
│ (Concatenate) │ 128) │ │ sequential_2[0][… │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ conv2d_transpose_8 │ (None, 256, 256, │ 6,147 │ concatenate_6[0]… │
│ (Conv2DTranspose) │ 3) │ │ │
└─────────────────────┴───────────────────┴────────────┴───────────────────┘
Total params: 54,425,859 (207.62 MB)
Trainable params: 54,414,979 (207.58 MB)
Non-trainable params: 10,880 (42.50 KB)
Our final loss values look like this:
Generator Total Loss: 1.0904
Generator GAN Loss: 1.0902
Generator L1 Loss: 0.0000
Discriminator Loss: 1.0119
We used this model to perform classifications of our imagery from 2017 to 2024 using our customary tile based approach. We use OpenCV to detect features on each tile and smoothen the background and foreground separately to create a single grid structure of almost cartoonishly, but very immediately visible features; this makes for a much smaller image that can be opened on pretty much any device regardless of processing power while being immediately and intuitively understandable.
In this image, bright green is forest; untouched forestry. Blue is water. Various shades of green turning to brown shows green areas that have been cut down or trimmed; white of course represents built up areas. There is a difference with the bluescale imagery in that this classifies human activity and not natural features that happen to have very low amounts of greenery, for example, scrubland.
This gives us a handy grid structure that makes it easier to detect broad, sweeping changes before examining those precise locations with the more fine-grained imagery above, across the years from 2017-2024, with the grid aligned uniformly.
We (Team Watchdog) hope that this work will be useful to journalists, climate analysts, and various people interested in the changing shape and nature of Sri Lanka. These images and outputs are provided on the terms of the MIT license -
PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THESE OUTPUTS OR THE USE OR OTHER DEALINGS IN THE OUTPUTS.
This work was created by Watchdog Sri Lanka (Appendix) as a part of the project Building Tools to Strengthen Pluralist, Inclusive and Fact-based Public Discourse, conducted by LIRNEasia. LIRNEasia (www.lirneasia.net) is a pro-poor, pro-market regional digital policy think tank. The project is conducted in partnership with the Strengthening Social Cohesion and Peace in Sri Lanka (SCOPE) programme, co-funded by the European Union and German Federal Foreign Office. SCOPE is implemented by GIZ in partnership with the Ministry of Justice, Prisons Affairs and Constitutional Reforms.
This project contains modified Copernicus Sentinel data [2024] via Google Earth Engine and Sepal.io. EU law grants free access to Copernicus Sentinel Data and Service Information for the purpose of the following use in so far as it is lawful:
- reproduction;
- distribution;
- communication to the public;
- adaptation, modification and combination with other data and information;
- any combination of points (a) to (d)
Special thanks to the European Space Agency, Google and Sepal for making these services available.