WeatherAI 

Industrial organisations involved

WATERVIEW blends expertise in meteorology, hydrology, environmental modeling, computer vision and AI, looking for new solutions to solve challenging problems.

WaterView combines its cutting-edge computer vision technologies with simple cameras, to turn each cam into a non-dedicated multivariable sensor.

That is one small step for a camera, one giant leap to reduce the costs of granular, real-time weather and environmental monitoring. 

Technical/scientific Challenge:

The first generation of models are based on physical models, require both device calibration and installation calibration, trained on large, annotated dataset. In addition, these models require great hardware and software development.  

WeatherAI wants to develop a second generation of AI analytics, more portable and scalable than the first-generation ones, trained on optimized datasets. 

The Solution

GPUs were used for the model training. Parallel CPUs were used for image augmentation. Hyperparameters and datasets were optimized for 5 models. 

To complete the project, a TensorFlow 2.X library was developed, to simplify training, exporting and quantization of AI models for edge applications (https://github.com/waterviewsrl/swiss-army-keras). 

Randomized image augmentation pipelines, parallelized on node CPUs, were realized using the Albumentations library. 

To track dataset evolution using standard versioning tools (GIT), the Data Version Control System DVC was adopted. This helped the integration of Machine Learning / Data Science procedures and tools into the existing DevOps pipelines. 

Business impact:

The access to HPC resources helped in the development of the second generation of weather analytics, based on AI technologies. 

The solutions provide telemetries related to the current weather conditions (rain precipitations, visibility distance), as well as alerts related to specific dangerous conditions (flooding or snow settling on road pavement): 

  • RainCAM: detects the presence of rain precipitations, according to three severity classes 
  • VisCAM: measures the visibility distance, in the range 25 to 1500 meters 
  • FloodCAM: detects flooding events 
  • SnowCAM: detects snow settling on the ground 
  • SmoCAM: detects the presence of outdoor and indoor smoke and fire 

The benefits

Images Courtesy: WaterView