Daffodil helps an urban planning firm to leverage AI and detect green space from geo-spatial data 

The project began with our business analysts and software architects outlining the optimal architecture of the solution. They refined the functional requirements and developed a complete product vision and its development roadmap. Once the strategic plans were mutually finalized, Daffodil’s team mapped the requirements onto the technology landscape and suggested technologies such as QGIS, Leaflet, Maply, WhirlyGlobe, Node.js, AWS, TensorFlow, and GBDX as the core technologies for the solution development.

In order to identify the location of urban trees, team Daffodil opted for object detection techniques. Object detection is a computer vision technique for locating instances of objects in images or videos.

For building the data set for machine learning models, the spatial data captured through satellites were used. The data was stored in GeoTIFF files from which different band information was to be derived. Since different species of trees are identified in different bands, building a single ML model with numerous combinations couldn’t be a solution. Thus, as a solution to the problem, multiple ML models were developed and ensemble modeling was used for prediction.

Ensemble modeling is a process where multiple diverse models are created to predict an outcome, either by using many different modeling algorithms or using different training data sets. The ensemble model then aggregates the prediction of each base model and results in a once final prediction for the unseen data.

Team Daffodil discovered the trees that could be identified in a specific band. Using a combination of 3 bands in a model, 11 machine learning models were created to identify the tree species and their location. These models were further assembled for prediction in the Amsterdam and Boston region. More than 90% of trees, captured from different angles and in various seasons were identified using the models. 

Along with this, Green City Watch had the requirement to make the solution scalable. Since the reflection in the bands varies according to the species, a similar solution won’t work in different geographies. Therefore, for different geographies, a different band combination would be required.

In order to optimize the resources and cost involved in the development, a number of cloud services (AWS) were utilized. AWS SES for email notification, AWS ECR to store docker images, AWS Sagemaker to run docker containers in GPU environment, AWS S3 Bucket for data storage, AWS Lambda to scale up or down the resources, when required.