Regresar

Band-selection in spectroscopy information for vegetation health monitoring using explainable models

Abstract:

Health monitoring of vegetation can be achieved through the use of hyperspectral cameras and spectrometers. However, both types of equipment are expensive, and in the case of spectrometers, the data collection process is time- and resource-consuming. Even though the information from these instruments is comprehensive, for specific tasks, it is often redundant, with only a portion of the bands providing useful information for applications such as leaf water prediction. In this context, we propose the identification of such bands using the GRAD-CAM algorithm and 1D convolutional neural networks for predicting leaf water content. The input data consists of the spectral signatures of avocado, olive, and grape leaves in the range of 350 to 2500 nm at different drying stages. To evaluate the identified set of bands, various vegetation indices were predicted using convolutional neural networks and the bands selected by the GRAD-CAM algorithm. The results demonstrate that the prediction of vegetation indices can achieve a correlation factor of up to 0.94 in some cases within the testing dataset.