Plant drought impact detection using ultra-high spatial resolution hyperspectral images and machine learning

Abstract

Early drought stress detection is crucial for restoring productivity, ensuring recovery, and providing vital information for mortality prevention. Hyperspectral remote sensing which is sensitive to subtle changes in leaf constituents and canopy structure is widely used to study how drought stress manifests and evolves over time. Drought impacts are complicated, affecting multiple leaf properties at different stages of severity. Spectral indices are widely used for stress monitoring, given their sensitivity to one or more plant biochemical, biophysical, and structural properties. However, appropriate spectral bands are not always apparent for early drought detection and stress progression monitoring since drought response varies by species, from tolerators to competitors. A tolerator can produce secondary compounds to enhance leaf survival while a competitor allows stressed leaves to perish so that new growth can occur. Also, spectral indices typically utilize a few bands and therefore do not exploit information in other spectral regions that are also influenced by drought. With the development of machine learning and increased computational power, the use of full spectra and derivative spectra becomes possible, holding great potential to offer new insights into the drought impacts. In this study, full spectra and derivative spectra were integrated with advanced machine learning algorithms for early drought detection and classification. The results from these models were compared to those from traditional spectral index-based methods. The models were tested using drought-stressed smooth brome (Bromus inermis) grass grown in a chamber with 0% and 25% field precipitation, using close-range hyperspectral images. The results showed that the use of derivative spectra and the deep learning model allowed accurate drought detection with an overall accuracy of up to 97.5% and 100% while spectral indices and spectral index models either failed to detect early stress or could detect early stress but with much lower accuracy. This study demonstrated the usefulness of integrating deep learning and hyperspectral data in plant stress research for stress impact assessment, recovery, and mitigation.