SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction

arXiv (Cornell University), Journal Year: 2021, Volume and Issue: unknown

Published: Jan. 1, 2021

One unique property of time series is that the temporal relations are largely preserved after downsampling into two sub-sequences. By taking advantage this property, we propose a novel neural network architecture conducts sample convolution and interaction for modeling forecasting, named SCINet. Specifically, SCINet recursive downsample-convolve-interact architecture. In each layer, use multiple convolutional filters to extract distinct yet valuable features from downsampled sub-sequences or features. combining these rich aggregated resolutions, effectively models with complex dynamics. Experimental results show achieves significant forecasting accuracy improvements over both existing Transformer-based solutions across various real-world datasets. Our codes data available at https://github.com/cure-lab/SCINet.

Language: Английский

---

A roadmap towards predicting species interaction networks (across space and time)

Philosophical Transactions of the Royal Society B Biological Sciences, Journal Year: 2021, Volume and Issue: 376(1837), P. 20210063 - 20210063

Published: Sept. 20, 2021

Networks of species interactions underpin numerous ecosystem processes, but comprehensively sampling these is difficult. Interactions intrinsically vary across space and time, given the number that compose ecological communities, it can be tough to distinguish between a true negative (where two never interact) from false have not been observed interacting even though they actually do). Assessing likelihood an imperative for several fields ecology. This means predict species-and describe structure, variation, change networks form-we need rely on modelling tools. Here, we provide proof-of-concept, where show how simple neural network model makes accurate predictions about limited data. We then assess challenges opportunities associated with improving interaction predictions, conceptual roadmap forward towards predictive models explicitly spatial temporal. conclude brief primer relevant methods tools needed start building models, which hope will guide this research programme forward. article part theme issue 'Infectious disease macroecology: parasite diversity dynamics globe'.

Language: Английский

---

The power of forecasts to advance ecological theory

Methods in Ecology and Evolution, Journal Year: 2022, Volume and Issue: 14(3), P. 746 - 756

Published: Aug. 11, 2022

Abstract Ecological forecasting provides a powerful set of methods for predicting short‐ and long‐term change in living systems. Forecasts are now widely produced, enabling proactive management many applied ecological problems. However, despite numerous calls an increased emphasis on prediction ecology, the potential to accelerate theory development remains underrealized. Here, we provide conceptual framework describing how forecasts can energize advance theory. We emphasize opportunities future progress this area through forecast development, comparison synthesis. Our describes approach shed new light existing theories while also allowing researchers address novel questions. Through rigorous repeated testing hypotheses, help refine understand their generality across Meanwhile, synthesizing allows about relative predictability variables horizons scales. envision where is integrated as part toolset used fundamental ecology. By outlining relevance theory, aim decrease barriers entry broaden community using insight.

Language: Английский

---

MASTREE+: Time‐series of plant reproductive effort from six continents

Global Change Biology, Journal Year: 2022, Volume and Issue: 28(9), P. 3066 - 3082

Published: Feb. 16, 2022

Significant gaps remain in understanding the response of plant reproduction to environmental change. This is partly because measuring long-lived plants requires direct observation over many years and such datasets have rarely been made publicly available. Here we introduce MASTREE+, a data set that collates reproductive time-series from across globe makes these freely available community. MASTREE+ includes 73,828 georeferenced observations annual (e.g. seed fruit counts) perennial populations worldwide. These consist 5971 population-level 974 species 66 countries. The mean median length 12.4 10 respectively, 1122 series extend at least two decades (≥20 observations). For subset well-studied species, extensive replication geographical climatic gradients. describe open-access set, as a.csv file, an associated web-based app for exploration. will provide basis improved Additionally, enable investigation ecology evolution strategies plants, role driver ecosystem dynamics.Aún existen importantes vacíos en la comprensión de respuesta reproductiva las plantas al cambio medioambiental, parte, porque su monitoreo especies longevas requiere una observación directa durante muchos años, y estos conjuntos datos rara vez han estado disponibles. Aquí presentamos MASTREE +, base que recopila tiempo reproducción todo el planeta, poniendo disposición libre acceso para comunidad científica. + incluye 73.828 puntos anual georreferenciados (ej. conteos semillas frutos) poblaciones perennes mundo. Estas observaciones consisten temporales nivel población provenientes países. La mediana duración es años (media = años) conjunto 1.122 menos dos décadas observaciones). Para un subconjunto bien estudiadas, +incluye amplio replicadas gradientes geográficos climáticos. Describimos abierto disponible como archivo.csv aplicación web asociada exploración datos. proporcionará mejorar sobre medioambiental. Además, facilitará los avances investigación ecología evolución estrategias reproductivas papel vegetal determinante dinámica ecosistemas.

---

Progress and opportunities in advancing near‐term forecasting of freshwater quality

Global Change Biology, Journal Year: 2023, Volume and Issue: 29(7), P. 1691 - 1714

Published: Jan. 9, 2023

Abstract Near‐term freshwater forecasts, defined as sub‐daily to decadal future predictions of a variable with quantified uncertainty, are urgently needed improve water quality management ecosystems exhibit greater variability due global change. Shifting baselines in land use and climate change prevent managers from relying on historical averages for predicting conditions, necessitating near‐term forecasts mitigate risks human health safety (e.g., flash floods, harmful algal blooms) ecosystem services water‐related recreation tourism). To assess the current state forecasting identify opportunities progress, we synthesized papers published past 5 years. We found that is currently dominated by quantity fewer number early stages development (i.e., non‐operational) despite their potential important preemptive decision support tools. contend more critically poised make substantial advances based examples recent progress methodology, workflows, end‐user engagement. For example, systems can predict temperature, dissolved oxygen, bloom/toxin events days ahead reasonable accuracy. Continued will be greatly accelerated adapting tools approaches machine learning modeling methods). In addition, effective operational require substantive engagement end users throughout forecast process, funding, training opportunities. Looking ahead, provides hopeful face increased risk change, encourage scientific community incorporate research management.

Language: Английский

---

Algal bloom ties: Systemic biogeochemical stress and Chlorophyll-a shift forecasting

Ecological Indicators, Journal Year: 2023, Volume and Issue: 154, P. 110760 - 110760

Published: Aug. 12, 2023

Algae are deemed to be the highest carbon drawdown contributors in world. Almost half of world's organic is fixed by marine phytoplankton, especially during algal blooms, despite accounting for less than one percent total photosynthetic biomass on Earth. Nonetheless, there growing concerns about rise harmful blooms terms their persistence and distribution that a fingerprint both global climate local anthropogenic pressure. A novel forecasting model – combining transfer entropy network inference convolution neural-network used predict bloom non-bloom epidemic regimes, environmental determinants define sources, causes systemic risk. The shows high skills extracting salient ecosystem features even without spatial dependencies. We defined 2D entropic mandala where ecological impact, manifested distribution's randomness Cyanobacteria-driven chlorophyll-a (CHL-a), proportional pressure/stress determined set erratic ocean/climate coastal/nutrient factors. Originally, risk was based CHL-a magnitude, shifts. Considering temporal variability Florida Bay (FL Bay) as case-study, we show how shifts becoming more persistent shallow areas with higher dinoflagellate/diatom ratio. This emphasizes likely key role cyanobacteria disorganization into phytoplankton organization, thus land discharge microbiome balance. presents major challenges considering increasing potential causality green–blue (river-dominated) red-tides cascading socio-ecological effects, including cycling alteration ingrained eutrophication coastal ecosystems. universal threshold top 20% Pareto extremes clearly defines regimes independently being endemic or epidemic, underpinned distinct eco-environmental interactions, largest biogeochemical stress structurally scale-free hub. Forecasting short- long-term indispensable quantifying health coastal-marine habitats, species humans, well impacts processes such sequestration capacity. However, decreases our ability except outbreaks when too late prevention. has investigating controlling undesired emergence spread.

Language: Английский

---

On the forecastability of food insecurity

Scientific Reports, Journal Year: 2023, Volume and Issue: 13(1)

Published: March 16, 2023

Abstract Food insecurity, defined as the lack of physical or economic access to safe, nutritious and sufficient food, remains one main challenges included in 2030 Agenda for Sustainable Development. Near real-time data on food insecurity situation collected by international organizations such World Programme can be crucial monitor forecast time trends insufficient consumption levels countries at risk. Here, using observations combination with secondary conflict, extreme weather events shocks, we build a forecasting model based gradient boosted regression trees create predictions evolution up 30 days future 6 (Burkina Faso, Cameroon, Mali, Nigeria, Syria Yemen). Results show that number available historical is key element performance. Among studied this work, those longest series, Yemen, proposed allows prevalence people into higher accuracy than naive approach last measured only. The framework developed work could provide decision makers tool assess how will evolve near clearly point added value continuous collection sub-national level.

Language: Английский

---

Evolutionary ecology of masting: mechanisms, models, and climate change

Trends in Ecology & Evolution, Journal Year: 2024, Volume and Issue: 39(9), P. 851 - 862

Published: June 10, 2024

Many perennial plants show mast seeding, characterized by synchronous and highly variable reproduction across years. We propose a general model of masting, integrating proximate factors (environmental variation, weather cues, resource budgets) with ultimate drivers (predator satiation pollination efficiency). This shows how the relationships between masting shape diverse responses species to climate warming, ranging from no change lower interannual variation or reproductive failure. The role environmental prediction as driver is being reassessed; future studies need estimate accuracy benefits acquired. Since central plant adaptation change, understanding adapts shifting conditions now question.

Language: Английский

---

Wavelet entropy-based evaluation of intrinsic predictability of time series

Chaos An Interdisciplinary Journal of Nonlinear Science, Journal Year: 2020, Volume and Issue: 30(3)

Published: March 1, 2020

Intrinsic predictability is imperative to quantify inherent information contained in a time series and assists evaluating the performance of different forecasting methods get best possible prediction. Model measure probability success. Nevertheless, model or does not provide understanding for improvement Intuitively, intrinsic delivers highest level informative unfolding whether system unpredictable chosen poor choice. We introduce novel measure, Wavelet Entropy Energy Measure (WEEM), based on wavelet transformation entropy quantification series. To investigate efficiency reliability proposed forecast was evaluated via networks approach. The uses energy distribution at scales compares it with white noise as deterministic random. test WEEM using wide variety ranging from deterministic, non-stationary, ones contaminated noise-signal ratios. Furthermore, relationship developed between Nash–Sutcliffe Efficiency, one widely known measures performance. demonstrated by exploring logistic map real-world data.

Language: Английский

---

Advances in forecasting harmful algal blooms using machine learning models: A case study with Planktothrix rubescens in Lake Geneva

Harmful Algae, Journal Year: 2020, Volume and Issue: 99, P. 101906 - 101906

Published: Sept. 29, 2020

Language: Английский

---