MACHINE LEARNING-BASED ESTIMATION OF CANOPY CHLOROPHYLL CONTENT IN CROPS FROM MULTIPLE SATELLITE IMAGES WITH VARIOUS SPATIAL RESOLUTIONS

The capability of quantifying canopy chlorophyll content from multiple satellite images with various spatial resolutions using machine learning approaches has been explored. The performance of several machine learning models has been assessed for Landsat-7, RapidEye, and PlanetScope satellite imagery, all acquired in 2017 over the Kellogg Biological Station (KBS), the largest off-campus research field station of Michigan State University. Retrieval uncertainty of the models has been characterized in relation to different spatial resolutions of the imagery and associated spectral bands. The preliminary results demonstrated that the best performance was achieved by the Gaussian process regression (GPR) method for RapidEye, with the coefficient of determination of R²=0.95, while neural network performed relatively well for Landsat 7 (R²=0.75), and GPR had best but somewhat lower performance for PlanetScope (R²=0.68). The best performance with RapidEye suggested that its red edge spectral band overpowered the capability of PlanetScope’s finer spatial resolution. Four different agriculture treatments used to grow the corn crop have also been incorporated in the analysis to quantify the retrieval of canopy chlorophyll content for each treatment. Data fusion has been applied to quantify advantages and disadvantages and to bridge the gap between different properties of the imagery.