Daffodil helps an urban planning firm to leverage AI and detect surface water bodies from geo-spatial data

To start with, team Daffodil worked on creating a data set that is comprehensive and reliable to provide a high accuracy rate. Moreover, the most suitable technique for model training was identified that’s capable of accurate predictions for existing, new, or unforeseen data. 

For model training, the image segmentation technique was utilized to identify water bodies and measure their stress levels. Image segmentation is one of several fields in Machine Learning where a model is trained in such a way that it can identify objects at the pixel level and classify it. This helps to accurately extract the objects of interest from the image.

Model Training: 

For training the models, the semantic segmentation technique was utilized. It refers to the task of assigning a class label to every pixel in the image. This way, a model treats multiple objects from a single class as a single entity. 

Data Set: 

For the data set, Green City Watch had Satellite imagery from World View 1 in the TIFF format. Since the TIFF are detailed and has a lot of metadata stored, it’s usually large in size. Thus, the images were converted into JPEG format (which is less complex and has only 3 bands and spectral indices). 

Moreover, several measures were taken to remove noise from the data set. Atmospheric correction (such as removing excessive clouds from imagery), identifying gaseous components, converting large-sized images into grids, etc.

Water Boundaries Detection:

For detecting the boundaries of the water bodies, Normalized Difference Water Index (NDWI) was calculated. It differentiates the water from the dry land and is most suitable for water body mapping. An ML model with a given and augmented data set was developed with 95% accuracy in detecting the surface water bodies.