About

Katia Fernandes is a Lecturer and Associate Research Scientist at the International Research Institute for Climate and Society, part of The Earth Institute at Columbia University. Her work focuses on hydrometeorology and climate variability, with a particular emphasis on the Amazon region. Her thesis, titled “The Amazon hydrometeorology and climate in ERA40,” explored climate variability and change, hydrometeorology, and the interactions between tropical and extra-tropical climates. At the IRI, Fernandes collaborates with Walter Baethgen and Lisa Goddard to investigate and develop plausible scenarios of regional climate change over the Amazon, including changes in seasonality for decadal to multi-decadal time horizons. Her research aims to examine patterns of fire use, spread, control, and losses due to uncontrolled fires, applying climate scenarios to better understand these phenomena.

Research topics

• Geography
• Mathematics
• Statistics
• Climatology
• Meteorology
• Medicine
• Environmental science
• Ecology
• Cartography
• Geology

Selected publications

• Evaluating the impacts of climate variability and change versus methodological approaches on stormwater control measures rainfall thresholds. A case study from North Carolina, USA

  2026-03-13

  articleOpen access1st authorCorresponding

  Climate change can significantly affect the effectiveness and resilience of Stormwater Control Measures (SCMs) in urban environments, as changes in rainfall intensity and depth can impact SCMs' capacity to mitigate flooding and reduce pollutant loads in water bodies. Thus, increasingly adaptive and robust SCM designs need to be informed by a continued understanding of changing precipitation patterns. ​In this study we evaluate SCMs precipitation thresholds based on two main methods: 90% rainfall capture depth and 90th percentile rank of daily precipitation over the period 1980-2023 in North Carolina, USA. We sought to address the questions of whether changes in daily precipitation are detected over time and, if so, whether these changes result in different thresholds depending on the method of choice. Our results indicate over the entire timeseries (1980-2023) both methods result in thresholds consistent with the current North Carolina Department of Environmental Quality SCM standard of 25.4 mm/day in central and western North Carolina and 38.1 mm/day in the eastern coastal plains. However, when the data is sliced in four 11-year periods the evolution of precipitation thresholds show a positive trend in both methods. We also find that the capture depth method is considerably more sensitive to extreme precipitation events than the 90th percentile method. Our results indicate that current water quality event standards in North Carolina may underestimate pollutant load treatment due to observed precipitation changes in recent years, suggesting a need for decadal adjustments. ​Defined SCM threshold need also to account for differences arising from the choice of calculation method, and practitioners using the capture depth methodology in particular, may want to revisit established thresholds.

  PublisherDOI

• Probabilistic forecasting of multiple air pollutants via latent dynamics modelling with deep learning

  2026-03-14

  articleOpen accessCorresponding

  Fine particulate matter (PM2.5), ozone (O3) and nitrogen dioxide (NO2) each pose significant risks to public health and are among the World Health Organisation (WHO) criteria pollutants. Operational air quality forecasting relies on computationally expensive Chemical Transport Models (CTMs), and recent deep learning methods focus on station-based forecasts, limiting usability to areas with station networks. We present a deep learning framework for probabilistic, gridded ambient air pollution forecasting to address both limitations. Our approach employs a latent dynamics architecture. A convolutional variational autoencoder (Conv-VAE) learns compressed latent representations of input channels. A temporal core block captures the dynamical evolution of ambient pollutants in the latent space, and a probabilistic decoder reconstructs forecasts with uncertainty intervals. Probabilistic forecasting allows for more trustworthy predictions, as stakeholders are also presented with relevant confidence. We systematically compare four latent cores: ConvLSTM, Mamba (state-space model), Transformer (attention) and Neural ODEs. This comparison will identify which inductive bias best represents the dynamics of ambient air pollution evolution. Experiments utilise a dataset of CAMS European reanalysis and ERA5 reanalysis (by ECMWF), as well as EDGAR emissions inventories over the UK (2015-2022), targeting 24–72 hour forecast horizons. Multi-pollutant settings test the framework's capacity to represent species with distinct atmospheric and chemical interactions in a unified latent representation. We will evaluate forecast skill, uncertainty quantification and computational efficiency of all models. Ongoing work is exploring physics-informed constraints, stochastic latent dynamics, and self-supervised pre-training for improved generalisation.

  PublisherDOI

• Unveiling the hydrologic climate resilience of Blue Green Infrastructure: Do we have our design/modelling numbers right?

  OJS-Service ULBT · 2026-03-23

  articleOpen access

  As precipitation patterns change, so does the need for revised Blue-Green Infrastructure (BGI) design standards. Holistic hydrologic design and modelling are vital for addressing climate uncertainties and ensuring the long-term integrity of optimized BGI. Within that scope, two key design needs are what size storm event needs to be safely routed through a BGI practice, and what is the precipitation depth of a Water Quality (WQ) event? WQ events principally determine the surface area and storage depths of BGI. The investigation uses North Carolina, USA, as a case study to determine revised design standards for these hydrologic parameters. Of particular interest was the increase, usually substantially (up to 25%), required for the WQ event depth when focusing on the most recent decade of available precipitation data. These results will be used to inform design standards in NC, likely yielding larger, and more protective, BGI in the coming years.

  PublisherDOI

• Machine Learning and Statistical Modeling of Air Pollution and Hospitalizations in South America’s Largest Metropolitan Area

  2026-03-14

  articleOpen accessCorresponding

  Air pollution is one of the main environmental and public health challenges in urban and rural areas, influenced by a wide range of factors, including traffic, biomass burning, and meteorology. In Brazil, about 326,478 deaths occurred between 2019 and 2021 due to exposure to air pollution. About 8,400 deaths per year are attributed to the Metropolitan Area of São Paulo (MASP), the largest metropolitan area of South America. Mitigating the effects of air pollution is only possible with a deep understanding of the spatial and temporal distributions of air pollutants at high resolution. We employed a machine learning framework based on Extreme Gradient Boosting (XGBoost) to spatialize particulate matter concentrations (PM2.5 and PM10) at MASP at 300 × 300 m². In addition, we developed a Ridge regression model to control multicollinearity and ensure stable estimates. We used this model to examine monthly hospitalizations associated with air pollution and heat exposure in MASP during 2023–2024, a period marked by severe biomass burning and heat waves. The study used integrated data from the Environmental Company of the State of São Paulo (CETESB), ERA5 reanalysis, land use and land cover (MapBiomas), emission inventories, terrain roughness and altitude, and hospitalizations (National Health Data Network, DATASUS) from 2022 to 2024. The XGBoost model has shown to be robust, with high R² values of 0.85 for PM2.5 and 0.88 for PM10, and RMSE of 3.3 µg/m³ and 5.2 µg/m³, respectively, for the test set (30% of the data). The analysis showed higher pollution levels in densely populated and industrialized areas, such as Guarulhos-Pimentas and Parque Don Pedro, while less urbanized regions, such as Pico do Jaraguá, had lower concentrations due to meteorological and topographical factors. The Ridge distributed-lag hospitalization model exhibited high explanatory power (R² = 0.88; RMSE = 214 hospitalizations per month). Chronic cumulative exposure over three months revealed that ozone and nitrogen dioxide were the dominant drivers of hospitalizations, associated with increases of approximately 65% and 57%, respectively, in monthly hospitalizations, while PM10 showed a moderate effect (~16%). Carbon monoxide did not present a significant association. These findings indicate that photochemical pollution combined with seasonal and thermal variability plays a critical role in respiratory morbidity in MASP, providing a robust quantitative basis for environmental health surveillance and urban air-quality management.

  PublisherDOI

• Socioeconomic and land use drivers of weedy grass infestation in the Peruvian Amazon

  Regional Environmental Change · 2025-11-26

  article
  PublisherDOI

• Landcover-categorized fires respond distinctly to precipitation anomalies in the South-Central United States

  Frontiers in Environmental Science · 2024-08-06 · 1 citations

  articleOpen access1st authorCorresponding

  Satellite detection of active fires has contributed to advance our understanding of fire ecology, fire and climate dynamics, fire emissions, and how to better manage the use of fires as a tool. In this study, we use active fire data of 12 years (2012–2023) combined with landcover information in the South-Central United States to derive a monthly, open-access dataset of categorized fires. This is done by calculating a fire predominance index used to rank fire-prone landcovers, which are then grouped into four main landscapes: grassland, forest, wildland, and crop fires. County-level aggregated analyses reveal spatial distributions, climatologies, and peak fire months that are particular to each fire type. Using the Standardized Precipitation Index (SPI), it was found that during the climatological fire peak-month, the SPI and fires exhibit an inverse relationship in forests and crops, whereas grassland and wildland fires show less consistent inverse or even direct relationship with the SPI. This varied behavior is discussed in the context of landscapes’ responses to anomalies in precipitation and fire management practices, such as prescribed fires and crop residue burning. In a case study of Osage County (OK), we find that large wildfires, known to be closely related to climate anomalies, occur where forest fires are located in the county and absent in areas of grassland fires. Weaker grassland fire response to precipitation anomalies can be attributed to the use of prescribed burning, which is normally planned under environmental conditions that facilitate control and thus avoided during droughts. Crop fires, on the other hand, are set to efficiently burn residue and are practiced more intensely in drier years than in wetter years, explaining the consistently strong inverse correlation between fires and precipitation anomalies. In our increasingly volatile climate, understanding how fires, vegetation, and precipitation interact has become imperative to prevent hazardous fire conflagrations and to better manage ecosystems.

  PublisherOA PDFDOI

• Subseasonal fire forecast in the Amazon using week-2 precipitation forecast combined with a vegetation health indicator

  2023 · 1 citations

  1st authorCorresponding
  • Environmental science
  • Meteorology
  • Climatology

  Numerical predictions for a lead time of 2 to 4 weeks, a timescale known as subseasonal, has only in recent years begun transitioning from research to operational settings. One experiment dedicated to that effort is the Subseasonal Experiment (SubX). In here, SubX multi-model ensemble (MME) mean precipitation forecast (2017-2021) for days 8 to 14 (week-2 forecast) is used as a covariate in logistic regression models to predict fire risk in the Amazon. The hybrid (dynamical and statistical) modelling approach describes the NextGen methodology aimed at improving forecast outcomes at the seasonal and subseasonal time scales. In a complementary experiment, a vegetation health index (VHI) is added to SubX precipitation forecasts as a predictor to fires. The findings show that fire risk can be skillfully assessed in most of the Amazon where fires occur regularly. In some sectors, SubX week-2 precipitation alone is a reliable predictor of fire risk, but the addition of VHI results both in (i) a larger portion of the Amazon domain with skillful forecasts and; (ii) higher skill in some sectors. The added information provided by VHI as a predictor is most relevant where the mosaic of land covers includes savannas and grassland, whereas SubX precipitation can be used as the sole predictor for week-2 fire risk forecast where the mosaic of land cover is dominated by forests. The operationalization of the methods presented in this study could allow for better preparedness and fire risk reduction in the Amazon with a lead time greater than a week.

  PublisherDOI

• The Sustainable Expansion of the Cocoa Crop in the State of Pará and Its Contribution to Altered Areas Recovery and Fire Reduction

  Journal of Geographic Information System · 2022-01-01 · 14 citations

  articleOpen access

  The state of Pará, located in the Amazon region of Brazil, has observed in recent years an increase in cocoa (Theobroma cacao) cultivation and has become the largest producer in Brazil. Due to its physiological characteristics, cacao is cultivated in native forests understory or under the shade produced by fast-growing native tree species, serving as an important species for restoration of degraded areas. However, mapping and monitoring cocoa plantation using optical sensor images is a challenge given its botanical and arboreal characteristics that can be confused with other native species at various stages of secondary regrowth. Agroforestry systems are important components of sustainable production in the Amazon and our work sought to better describe the evolution of cocoa plantations in terms of their historical expansion, farming properties practices, land use transitions and fire regimes. Our findings to analyze the relationships between cocoa plantations and hotspots, data from the INPE’s reference satellite between the years 2004 to 2020 were used in this study, polygons classified as cocoa areas, generated by the MapCacau research project, were used, in a total of 69,904 hectares distributed throughout the state of Pará. Finally, we used the protected areas’ official limits in the State of Pará to analyze the plantations’ occurrence in regions in discordance with environmental legislation. The data show that cocoa-producing properties are statistically fewer than non-producing properties, as well as having lower deforestation rates. In our study, we observed that 52,778 hectares (88.87%) of the cocoa area planted had already been deforested by the year 2008—the threshold of deforestation defined by Brazil’s Forest Code. It was also possible to verify that approximately 20,900 hectares continue to be mapped as forest by PRODES, despite our field data identifying cocoa plantations shaded by explored forest in these areas. Regarding the crop’s formation, the data show a tendency to convert pasture areas to cocoa plantations, proving that cocoa farming expansion in the State of Pará is an important activity for degraded areas recovery and not a main driver of deforestation. The finding that cocoa plantations are still classified as forest by PRODES and project TerraClass highlights the difficulty of mapping this crop using orbital images in a traditional way. Through this paper, it was possible to observe that due to the typical characteristics of perennial crops (cocoa), fire points showed a significant reduction in the mapped areas, highlighting that the expansion of cocoa plantations in the state of Pará contributed to soil protection, to the reduction of greenhouse gas emissions into, in addition to contributing to the generation of jobs and revenue. Finally, we found about 99.54% of the cacao plantations in the State of Pará are located outside of any preservation area, indigenous land or quilombola settlement.

  PublisherOA PDFDOI

• Combining precipitation forecasts and vegetation health to predict fire risk at subseasonal timescale in the Amazon

  Environmental Research Letters · 2022 · 9 citations

  1st authorCorresponding
  • Environmental science
  • Climatology
  • Meteorology

  Abstract Current forecast systems provide reliable deterministic forecasts at the scale of weather (1–7 days) and probabilistic outcomes at the scale of seasons (1–9 months). Only in recent years research has begun transitioning to operational settings to provide numerical predictions for a lead time of 2–4 weeks, a timescale known as subseasonal. The Subseasonal Experiment (SubX) multi-model ensemble mean precipitation forecast (2017–2021) for days 8–14 (week-2 forecast) is used as a covariate in logistic regression models to predict fire risk in the Amazon. In a complementary experiment, a vegetation health index (VHI) is added to SubX precipitation forecasts as a predictor of fires. We find that fire risk can be skillfully assessed in most of the Amazon where fires occur regularly. In some sectors, SubX week-2 precipitation alone is a reliable predictor of fire risk, but the addition of VHI as a predictor results both in (a) a larger portion of the Amazon domain with skillful forecasts and; (b) higher skill in some sectors. By comparing two sectors of the Amazon, we find that the added information provided by VHI is most relevant where the mosaic of land covers includes savannas and grassland, whereas SubX precipitation can be used as the sole predictor for week-2 fire risk forecast in areas where the mosaic of land cover is dominated by forests. Our results illustrate the potential for using numerical model forecasts, at the subseasonal timescale, in combination with satellite remote sensing of vegetation to obtain skillful fire risk forecasts in the Amazon. The operationalization of the methods presented in this study could allow for better preparedness and fire risk reduction in the Amazon with a lead time greater than a week.

  PublisherOA PDFDOI

• Detecting vulnerability of humid tropical forests to multiple stressors

  One Earth · 2021-07-01 · 101 citations

  articleOpen access
  PublisherOA PDFDOI

Frequent coauthors

• Miguel Pinedo-Vásquez

  37 shared

• Walter Baethgen

  23 shared

• Víctor Hugo Gutiérrez-Vélez

  Temple University

  23 shared

• María Uríarte

  Columbia University

  17 shared

• Ruth DeFries

  Columbia University

  13 shared

• Christine Padoch

  10 shared

• Louis Verchot

  10 shared

• Rong Fu

  University of California, Los Angeles

  9 shared

Labs

• Kátia Fernandes LabPI

Education

• PhD, Earth and Atmospheric Sciences

  Georgia Institute of Technology

  2009

• MSc, Meteorology

  National Institute for Space Research

  1996

• BSc, Meteorology

  Universidade Federal de Pelotas

  1993