Climate change is causing water scarcity in many arid regions of the world. Can we look to ancient technologies to help manage this dwindling resource? To explore this, I developed a machine learning model to detect the presence of ancient irrigation tunnels from satellite imagery. The model uses the Keras TensorFlow library to train and validate a convolutional neural network (CNN).

Karez or Qanat are ancient hand dug irrigation tunnels that have been used for thousands of years to transport groundwater over long distances in semi-arid regions. They are constructed by digging a gentle down sloping tunnel underground so that water can flow under gravity. The technique minimises water loss though evaporation and does not require any electrical pumping of ground water reserves, making the tunnels a sustainable source of water supply. Air flow over the wells can be used to cool houses and buildings. The tunnels are still used today for irrigation and domestic water supply in some regions of the world, although many have fallen into a state of disrepair. The tunnels are mostly found in countries along the old silk road, particularly in Afghanistan, Iran and Uzbekistan. Access to the tunnel is via a series of well-like vertical shafts which can be seen by satellite as a line of equally spaced circles.

To create training and test data, I visually selected Karez images using free 1m satellite imagery from the OrbView-3 commercial satellite data (from USGS). I then used the TensorFlow library to train and test the image classifier using 2960 training images and 740 validation images. To expand the size of the training data I augmented the images using the Keras ImageDataGenerator function.

A Jupyter notebook containing the code is stored here

The image classifier works well, detecting the presence of Karez tunnels with a precision of 95%. The next step is to productionise the model to run on multiple satellite tiles. This will enable mapping of the locations of tunnels across multiple countries in Asia.

Example training images of Karez shafts. Data comes from the OrbView-3 commercial satellite data and has a resolution of 1m resolution.

Left: Visual selection of training images using an OrbView-3 tile located in Naru Kalay in the West of Kandahar, Afghanistan. Karez in orange, not karez in green. Right: training and validation loss and accuracy.