Machine learning prediction of coffee rust severity on leaves using spectroradiometer data

Abstract

The interest in using remote sensing data in agriculture, including plant disease assessments, has increased considerably in the last years. The satellite-based Sentinel-2 MultiSpectral Imager (MSI) sensor has been launched recently for multispectral vegetation condition assessment for agricultural and ecosystem applications. The aim of this pilot study conducted in the greenhouse using a hand-held spectroradiometer was to assess the utility of the same wavebands as used in the Sentinel-2 MSI in assessing and modeling coffee leaf rust (CLR) based on the non-linear radial basis function-partial least squares regression (RBF-PLS) machine learning algorithm, compared with ordinary partial least squared regression (PLSR). The RBF-PLS derived models satisfactorily described CLR severity (R2=0.92 and RMSE=6.1% with all bands and R2=0.78 and RMSE=10.2% with selected bands) when compared with PLSR (R2 = 0.27 and RMSE = 18.7% with all bands and R2 = 0.17 and RMSE = 19.8% with selected bands). Specifically, four bands, B2 (490 nm), B4 (665 nm), B5 (705 nm) and B7 (783 nm) were identified as the most important spectral bands in assessing and modeling CLR severity. Better accuracy was obtained for most severe levels of CLR (R2=0.71 using all variables) than for moderate levels (R2=0.38 using all variables). Overall, the findings of this study showed that the use of RBF-PLS and the four Sentinel-2 MSI bands could enhance CLR severity estimation at the leaf level. Further work will be needed to extrapolate these findings to the crop level using the Sentinel-2 platform.