Mini-batching ecological data to improve ecosystem models with machine learning

Abstract

AbstractEcosystems are involved in global biogeochemical cycles that regulate climate and provide essential services to human societies. Mechanistic models are required to describe ecosystem dynamics and anticipate their response to anthropogenic pressure, but their adoption has been limited in practice because of issues with parameter identification and because of model inaccuracies. While observations could be used to directly estimate parameters and improve models, model nonlinearities as well as shallow, incomplete and noisy datasets complicate this process. Here, we propose a machine learning (ML) framework relying on a mini-batch method combined with automatic differentiation and state-of-the-art optimizers. By splitting the data into mini-batches with a short time horizon, we show both analytically and numerically that the mini-batch method regularizes the learning problem. When combined with the proposed numerical implementation, the resulting ML framework can efficiently learn the parameter of complex dynamical models and is a workhorse for model selection. We evaluate the performance of the ML framework in recovering the dynamics of a simulated food-web. We show that it can efficiently learn from noisy, incomplete and independent time series, accurately estimating the model parameters and providing reliable short-term forecasts. We further show that the ML framework can provide statistical support for the true generating model among several candidates. In summary, the proposed ML framework can efficiently learn from data and elucidate mechanistic pathways to improve our understanding and predictions of ecosystem dynamics.Author summaryEcosystem models which explicitly represent ecological mechanisms are required to forecast ecosystem responses to global changes, but large mismatches with observations limit their predictive ability. To help address this major problem, we propose a novel machine learning (ML) method aiming at improving ecosystem models with data. The ML method is based on a learning strategy where the model is matched against small chunks of data, called mini-batches, and it involves numerical techniques commonly used in the training of neural networks. By benchmarking the performance of the ML method with a challenging food-web model, we show that our approach is robust against noise and partial observations, can process and combine the information contained in independent datasets, and can provide statistical support for the most adequate model among several candidates. Our proposed method therefore accommodates the reality of ecological datasets and our partial knowledge of ecosystem processes. By efficiently blending data and ecological theory with state-of-the-art ML techniques, our work offers novel tools to improve our understanding and predictions of ecosystem dynamics.