MUMBAI, India, May 1 -- Intellectual Property India has published a patent application (202641049921 A) filed by Pokkunuri Pardha Saradhi; A S C S Sastry; and Koneru Lakshmaiah Education Foundation, Vaddeswaram, Andhra Pradesh, on April 20, for 'enhanced rainfall forecasting using oversampling and scaling approaches.'
Inventor(s) include Dr A. S. C. S. Sastry; Dr N Prabakaran; Dr A V Prabu; Dr Ch. Sreevardhan; M. S. M. Mounika; Kalyani Chandra; Vinay Kallagunta; and Dr P. Pardhasaradhi.
The application for the patent was published on May 1, under issue no. 18/2026.
According to the abstract released by the Intellectual Property India: "Forecasting rainfall accurately is an essential factor in supporting agricultural planning processes, mitigating impacts from natural disasters, and directing the policy decisions that promote sustainable practices. The purpose of this study is to develop an alternative data-driven framework for predicting rainfall amounts through training and evaluating many different machine learning algorithms by using historical observational data from the National Oceanic and Atmospheric Administration (NOAA). The types of input variables available for our prediction models include: (1) atmospheric pressure; (2) surface temperature; (3) dewpoint temperature; (4) relative humidity; (5) cloud amount; (6) solar period; (7) wind direction; (8) and windspeed. Prior to building the predictive models, we conducted data integrity checks on the dataset to ensure the reliability of some of the data was met. We replaced the missing information with an imputation method (e.g., mean/mode), eliminated data we deemed non-contributory, and converted categorical variables into numeric representations that could be understood by our prediction models. Another challenge in the dataset was skewed class distributions among collections of rainfall amount data. Thus, we utilized a data augmentation method to generate synthetic examples for the minority class collection of observations to help improve the model's prediction accuracy on detecting rainfall events. To ensure the predictive models' performance was equal across all input feature variables, we also utilized normalization techniques such as StandardScaler to generate the same scale representation of all input feature variables. Four different classifier algorithms were developed and evaluated: logistic regression, support vector machine (SVM), random forest (RF), and artificial neural networks (ANN). We evaluated the performance of the models by their accuracy as well as the overall quality of classification using confusion matrices and performance measures (precision, recall, F1-score). Logistic Regression obtained the following accuracies: 85% to 88%; SVM's accuracy was a bit higher: 88% to 90%. Random Forest showed the most generalization with an accuracy of 90% to 93%. And Neural Network produced the final results of 92% to 95% accuracy. To increase performance, we employed GridSearchCV to optimize each model's hyperparameters. This tool attempts to find the best combination of hyperparameters by testing different combinations of values. Cross-validation was also performed so we could verify the performance of the models objectively. The top three features that impacted whether it would rain were humidity/relative humidity, cloudiness/amount of clouds, and dew point. These features improved our ability to predict weather patterns. Proper data preparation (i.e., correcting imbalances, normalizing features) and using highly tuned machine learning models improved our ability to forecast precipitation accurately. These systems can be easily integrated into real-time forecasting programs as well as automated systems for making environmental decisions."
Disclaimer: Curated by HT Syndication.
