Using Machine Learning Techniques to Develop an Effective Weather Prediction Model from Ground-Based Cloud Images

Abstract

The measurement, analysis and forecasting of weather has several use cases in climate modelling, agriculture, and farming. Every year, millions of lives are affected by heavy rains and storms. Therefore, it is important to predict the weather before time to avoid any disaster. Different types of sensors are used to collect data from the environment and then analysis is performed for necessary forecasting. The conventional process requires deployment of several sensors and consequently does not remain very cost-effective for small-scale localized deployment scenarios, especially in farming and agriculture. In recent years, computer vision and AI have made significant progress to address similar problems. In this research, we investigate the potential of ground-based images towards the development of weather prediction models. This study focuses on the development of cost-effective methods for weather-prediction in a localized environment that can be deployed using off-the-shelf, low-cost cameras pointed towards the sky. A custom data set is developed by using images from Singapore Whole Sky IMaging CATegories (SWIMCAT) dataset, Kaggle and images obtained through web-scrapping. The conventional machine learning algorithms i.e., Stochastic Gradient Descent (SGD) Classifier, Support Vector Classifier (SVC) and Random Forest Classifier (RFC) achieve 66%, 76% and 72% accuracy by using Histogram of Oriented Gradients (HoG) features. The study uses a Convolutional Neural Network (CNN) model that achieves 95% accuracy. However, the comparison of training and testing accuracy and validation scores show that the model is over-fitting. To overcome this limitation, a second CNN model has been used that shows a better generalization. The presented dataset demonstrates better class balance compared to previous work, as many models trained on the SWIMCAT dataset overlook class-balance issues. Moreover, the proposed CNN model utilizes fewer layers to tackle over-fitting concerns while maintaining similar accuracy levels. This work can be used for the development of an embedded system or a low-cost IoT system interfaced with a ground camera for low-cost weather forecasting.