An Iterative Learning Framework for Multimodal Chlorophyll-a Estimation

Abstract

Precise monitoring of the chlorophyll type “a” (chl-a) concentration is critical in determining the level of production of oxygen and, consequently, the health conditions of inland aquatic ecosystems. This paper addresses two important issues in building precise and robust regression models for chl-a concentration from remote sensing data: the presence of multimodality in the sensor data distribution and the scarcity of information available to properly label the data. In order to effectively deal with the aforementioned issues, we propose an iterative learning framework (iterative transductive environmental modeling system) based on the principles of transductive learning that combines data-driven regression-based modeling with an iterative nonlinear classification process. The classification procedure, contingent on the maximum margin principle, generates data sets associated with each statistical modality. The classified data are labeled through a process of consecutive selection of the best candidate samples. Different selection mechanisms are discussed. The proposed method was applied in the empirical assessment of chl-a concentration from MEdium Resolution Imaging Spectrometer and Moderate Resolution Imaging Spectroradiometer satellite data and validated by in situ measurements in Lake Winnipeg in Manitoba, Canada.