LAUSR

We present a novel method using satellite and biogeochemical Argo (BGC-Argo) data to retrieve the 3-D structure of chlorophyll $a$ (Chla) in the northern Indian Ocean (NIO). The random forest (RF)-based method infers the vertical distribution of Chla using the near-surface and vertical features. The input variables can be divided into three categories: 1) near-surface features acquired by satellite products; 2) vertical physical properties obtained from temperature and salinity profiles collected by BGC-Argo floats; and 3) the temporal and spatial features, i.e., day of the year, longitude, and latitude. The RF-model is trained and evaluated using a large database including 9738 profiles of Chla and temperature-salinity properties measured by BGC-Argo floats from 2011 to 2021, with synchronous satellite-derived products. The retrieved Chla values and the validation dataset (including 1948 Chla profiles) agree fairly well, with $R^{2} = 0.962$ , root-mean-square error (RMSE) = 0.012, and mean absolute percent difference (MAPD) = 11.31%. The vertical Chla profile in the NIO retrieved from the RF-model is more accurate and robust compared to the operational Chla profile datasets derived from the neural network and numerical modeling. A major application of the RF-retrieved Chla profiles is to obtain the 3-D Chla structure with high vertical resolution. This will help to quantify phytoplankton productivity and carbon fluxes in the NIO more accurately. We expect that RF-model can be used to develop long-time series products to understand the variability of 3-D Chla in future climate change scenarios.