Location Agnostic Adaptive Rain Precipitation Prediction using Deep Learning

TL;DR

This work tackles the challenge of cross-location rain precipitation forecasting under distribution shifts and climate-change-driven pattern changes. It introduces a location-agnostic adaptive deep learning framework that transfers knowledge from a source city (Dhaka) to target cities (Paris, Los Angeles, Tokyo) using a two-component loss: $L_{src}$ and $L_{tgt}$, combined as $L_{total} = \lambda_{1}L_{src} + \lambda_{2}L_{tgt}$. The approach yields substantial MAE improvements on target domains (e.g., 43.51% for Paris, 5.09% for Los Angeles, 38.62% for Tokyo) and outperforms several traditional ensembling baselines, demonstrating strong cross-domain generalization. The dataset comprises NASA POWER weather records from 2003–2023 with 20 features, enabling robust evaluation. The results suggest practical impact for location-agnostic rainfall forecasting and motivate further work on feature-efficient and unsupervised adaptation strategies.

Abstract

Rain precipitation prediction is a challenging task as it depends on weather and meteorological features which vary from location to location. As a result, a prediction model that performs well at one location does not perform well at other locations due to the distribution shifts. In addition, due to global warming, the weather patterns are changing very rapidly year by year which creates the possibility of ineffectiveness of those models even at the same location as time passes. In our work, we have proposed an adaptive deep learning-based framework in order to provide a solution to the aforementioned challenges. Our method can generalize the model for the prediction of precipitation for any location where the methods without adaptation fail. Our method has shown 43.51%, 5.09%, and 38.62% improvement after adaptation using a deep neural network for predicting the precipitation of Paris, Los Angeles, and Tokyo, respectively.

Figures (2)

• Figure 1: The total workflow is shown in this figure. (a) The deep neural network that is used for training. We formulate our problem as a regression problem. Hence there is only one output node. (b) Training on the source side. $X_{src}$ is the source data (weather features) and $\hat{Y}_{src}$ is the ground truth data of the target value (precipitation). The training is guided by the mean square error (MSE) loss. (c) The adaptation on the target side is guided by MSE loss using both source and target data.
• Figure 2: Performance comparison on predicting precipitation of different methods by training them on the weather data of Dhaka city and using the pre-trained model for testing on Paris, Los Angeles, and Tokyo. ADB , GRB, RFR, SR, DWOA, and DWA indicate Adaboost, Gradient Boosting Regressor, Random Forest Regressor, Stacking Regressor, Deep learning WithOut Adaptation, and Deep learning With Adaptation, respectively. We see that Deep learning with adaptation outperforms all other methods by securing the lowest Mean Absolute percentage error.