Monitoring the spatio-temporal variations of surface chlorophyll-a concentration (Chl, a proxy of phytoplankton biomass) greatly benefited from the availability of continuous and global ocean color satellite measurements from 1997 onward. These two decades of satellite observations are however still too short to provide a comprehensive description of Chl variations at decadal to multi-decadal timescales. This paper investigates the ability of a machine learning approach (a non-linear statistical approach based on Support Vector Regression, hereafter SVR) to reconstruct global spatio-temporal Chl variations from selected surface oceanic and atmospheric physical parameters. With a limited training period (13 years), we first demonstrate that Chl variability from a 32-years global physical-biogeochemical simulation can generally be skillfully reproduced with a SVR using the model surface variables as input parameters. We then apply the SVR to reconstruct satellite Chl observations using the physical predictors from the above numerical model and show that the Chl reconstructed by this SVR more accurately reproduces some aspects of observed Chl variability and trends compared to the model simulation. This SVR is able to reproduce the main modes of interannual Chl variations depicted by satellite observations in most regions, including El Niño signature in the tropical Pacific and Indian Oceans. In stark contrast with the trends simulated by the biogeochemical model, it also accurately captures spatial patterns of Chl trends estimated by satellite data, with a Chl increase in most extratropical regions and a Chl decrease in the center of the subtropical gyres, although the amplitude of these trends are underestimated by half. Results from our SVR reconstruction over the entire period (1979–2010) also suggest that the Interdecadal Pacific Oscillation drives a significant part of decadal Chl variations in both the tropical Pacific and Indian Oceans. Overall, this study demonstrates that non-linear statistical reconstructions can be complementary tools to in situ and satellite observations as well as conventional physical-biogeochemical numerical simulations to reconstruct and investigate Chl decadal variability.

中文翻译：

---

使用非线性统计方法重建全球叶绿素 a 变化

监测表面叶绿素 a 浓度（Chl，浮游植物生物量的代表）的时空变化极大地受益于 1997 年以来连续和全球海洋颜色卫星测量的可用性。然而，这两个十年的卫星观测仍然太短，无法全面描述十年到多年时间尺度的 Chl 变化。本文研究了机器学习方法（一种基于支持向量回归的非线性统计方法，以下简称 SVR）从选定的表面海洋和大气物理参数重建全球时空 Chl 变化的能力。在有限的培训期间（13 年），我们首先证明，通常可以使用模型表面变量作为输入参数，通过 SVR 巧妙地再现来自 32 年全球物理-生物地球化学模拟的 Chl 变异性。然后，我们使用来自上述数值模型的物理预测器应用 SVR 来重建卫星 Chl 观测，并表明与模型模拟相比，由该 SVR 重建的 Chl 更准确地再现了观测到的 Chl 变异性和趋势的某些方面。该 SVR 能够再现大多数地区卫星观测所描绘的年际 Chl 变化的主要模式，包括热带太平洋和印度洋的厄尔尼诺现象。与生物地球化学模型模拟的趋势形成鲜明对比的是，它还准确捕捉了卫星数据估计的 Chl 趋势的空间格局，大多数温带地区的 Chl 增加，副热带环流中心的 Chl 减少，尽管这些趋势的幅度被低估了一半。我们在整个时期（1979-2010 年）的 SVR 重建结果也表明，年代际太平洋涛动驱动了热带太平洋和印度洋年代际 Chl 变化的重要部分。总体而言，这项研究表明，非线性统计重建可以作为原位和卫星观测以及常规物理-生物地球化学数值模拟的补充工具，以重建和研究 Chl 年代际变异性。我们在整个时期（1979-2010 年）的 SVR 重建结果也表明，年代际太平洋涛动驱动了热带太平洋和印度洋年代际 Chl 变化的重要部分。总体而言，这项研究表明，非线性统计重建可以作为原位和卫星观测以及常规物理-生物地球化学数值模拟的补充工具，以重建和研究 Chl 年代际变异性。我们在整个时期（1979-2010 年）的 SVR 重建结果也表明，年代际太平洋涛动驱动了热带太平洋和印度洋年代际 Chl 变化的重要部分。总体而言，这项研究表明，非线性统计重建可以作为原位和卫星观测以及常规物理-生物地球化学数值模拟的补充工具，以重建和研究 Chl 年代际变异性。