A machine learning based prediction system for the Indian Ocean Dipole

Predicting maximum temperatures over India 10-days ahead using machine learning models

In the months of March-June, India experiences high daytime temperatures (Tmax), which sometimes lead to heatwave-like conditions over India. In this study, 10 different machine learning models are evaluated for their ability to predict the daily Tmax anomalies 10 days ahead in the months of March-June. Several model experiments were carried out to identify an optimal model to predict daily Tmax anomalies over India. The results indicate that the AdaBoost regressor with Multi-layer Perceptron as the base estimator is an optimal model to predict the Tmax anomalies over India in the months of March-June. The optimal model predictions are benchmarked against 10-day persistence predictions and the predictions from the Climate Forecast System (CFS) reforecast. The results indicate that the machine learning model skill is higher than persistence and comparable to CFS reforecast 10-day predictions in April and May. In March and June, the machine learning models have low skill scores and perform no better than persistence. These results indicate that the machine learning models are promising tools to predict the surface air maximum temperature anomalies over India in April and May and can complement predictions from more sophisticated numerical models.

Multi-task machine learning improves multi-seasonal prediction of the Indian Ocean Dipole

As one of the most predominant interannual variabilities, the Indian Ocean Dipole (IOD) exerts great socio-economic impacts globally, especially on Asia, Africa, and Australia. While enormous efforts have been made since its discovery to improve both climate models and statistical methods for better prediction, current skills in IOD predictions are mostly limited up to three months ahead. Here, we challenge this long-standing problem using a multi-task deep learning model that we name MTL-NET. Hindcasts of the IOD events during the past four decades indicate that the MTL-NET can predict the IOD well up to 7-month ahead, outperforming most of world-class dynamical models used for comparison in this study. Moreover, the MTL-NET can help assess the importance of different predictors and correctly capture the nonlinear relationships between the IOD and predictors. Given its merits, the MTL-NET is demonstrated to be an efficient model for improved IOD prediction.

Exploring the potential of machine learning for simulations of urban ozone variability

Machine learning (ML) has emerged as a powerful technique in the Earth system science, nevertheless, its potential to model complex atmospheric chemistry remains largely unexplored. Here, we applied ML to simulate the variability in urban ozone (O₃) over Doon valley of the Himalaya. The ML model, trained with past variations in O₃ and meteorological conditions, successfully reproduced the independent O₃ data (r² ~ 0.7). Model performance is found to be similar when the variation in major precursors (CO and NO_x) were included in the model, instead of the meteorology. Further the inclusion of both precursors and meteorology improved the performance significantly (r² = 0.86) and the model could also capture the outliers, which are crucial for air quality assessments. We suggest that in absence of high-resolution measurements, ML modeling has profound implications for unraveling the feedback between pollution and meteorology in the fragile Himalayan ecosystem.

Deep Learning of Sea Surface Temperature Patterns to Identify Ocean Extremes

We performed an out-of-distribution (OOD) analysis of ∼12,000,000 semi-independent 128 × 128 pixel2 sea surface temperature (SST) regions, which we define as cutouts, from all nighttime granules in the MODIS R2019 Level-2 public dataset to discover the most complex or extreme phenomena at the ocean’s surface. Our algorithm (ULMO) is a probabilistic autoencoder (PAE), which combines two deep learning modules: (1) an autoencoder, trained on ∼150,000 random cutouts from 2010, to represent any input cutout with a 512-dimensional latent vector akin to a (non-linear) Empirical Orthogonal Function (EOF) analysis; and (2) a normalizing flow, which maps the autoencoder’s latent space distribution onto an isotropic Gaussian manifold. From the latter, we calculated a log-likelihood (LL) value for each cutout and defined outlier cutouts to be those in the lowest 0.1% of the distribution. These exhibit large gradients and patterns characteristic of a highly dynamic ocean surface, and many are located within larger complexes whose unique dynamics warrant future analysis. Without guidance, ULMO consistently locates the outliers where the major western boundary currents separate from the continental margin. Prompted by these results, we began the process of exploring the fundamental patterns learned by ULMO thereby identifying several compelling examples. Future work may find that algorithms such as ULMO hold significant potential/promise to learn and derive other, not-yet-identified behaviors in the ocean from the many archives of satellite-derived SST fields. We see no impediment to applying them to other large remote-sensing datasets for ocean science (e.g., SSH and ocean color).