The problem Frida solves

Every year, heavy rainfall continues to affect most districts of Nepal causing floods, triggering landslides leading to a huge number of deaths and heavy damage to property. According to Nepal Disaster Risk Reduction Portal, over 39 landslides and 16 floods were reported in just the first week of July 2021. According to the report, a total of 38 people died and 51 were injured within the week. More than 1250 people had been evacuated, 5100 displaced and a total of 790 houses were destroyed.

The amount of chaos created by landslides and floods in 3 months of monsoon is astronomical and flood/landslide alerts often don’t come in time. With advanced deep learning and statistical analysis, Frida aims to solve the early warning problem and much more. Frida uses Neural Networks to predict if there’s any chance of landslide or flood disaster within 7 days and 5km range of the user. We aim to reduce the calamities caused by landslides and floods with the help of this advanced and efficient, low-maintenance system. This will greatly help users for the three months of the monsoon but what about the rest of the nine months?

The goal of Frida is not only to act as an alert system but be a one-stop solution for all climate action needs. Frida scrapes the web/google news to provide a summary of all international news from portals like The New York Times, The Washington Post, and much more related to climate change topics. On top of that, Frida also scrapes Nepali news portals like OnlineKhabar, and Setopati to bring the users local news on climate change. Frida also scrapes online events manager social media handle meetups to filter out any upcoming climate change-related events near the user and if the user is feeling generous, Frida would also scrape fundraising campaigns related to climate change. On top of that, Frida would be the last application the user would need to get weather updates with a personalized message surrounding the weather forecast sent daily to the user.

Challenges we ran into

Problems we faced

1. Lack of data availability on landslides and weather data.
2. Ambiguous endpoints on Hydrology Nepal’s website.
3. Non-standard data entry on Nasa’s landslides data entry.
4. Obfuscated google news scraping via bs4.
5. PyTorch was too big to install on any servers for free so we couldn't deploy our Deep Learning model on production itself.

Our solutions

1. Had to create our own data by relating two completely different dataset with different format of addresses by using Harper’s distance relation to conclude if weather at certain areas will trigger landslides events at a certain distance from the weather’s coordinates; creating a dataset with 25k positive examples and 26k negative examples.
2. Had to use a headless browser to scrape google news data.
3. Wondered for several continuous hours trying to map between indices and rivers at Hydrology Nepal’s website.