Abiotic stress detection using spectral information for crop monitoring

Abstract

Remote sensing is one of the technologies with the potential for precision agriculture ap plications. Remote sensing systems include passive sensors, such as multispectral and hy perspectral sensors, which measure the energy reflected or emitted by a surface along the electromagnetic spectrum. Remote sensing allows monitoring large areas in less time than regular soil analysis processes. Several studies have demonstrated the potential of spectral data to crop stress conditions. However, most of these studies are limited to spectral signatu res taken in situ. Some works estimate crop conditions from multispectral and hyperspectral images, but most use vegetation indeces, which do not take full advantage of the spatial and spectral data captured by spectral cameras. Despite the continuing development of precision agriculture based on remote sensing, there is still ample scope for further studies to meet the agricultural sector’s needs. This thesis focuses on the extracting information from spectral data to detect crop stress conditions. The study was developed in two scales. The first one seeks the spectral characterization of stressed crops from spectral signatures collected in situ. The second one studies the capacities and limitations of remotely captured spectral imagery for stress detection, considering spatial information. This work developed a framework for water and nutritional stress detection using crop signatures combining the capabilities of either band ratios, discriminative bands, or the full spectra with supervised classifiers to detect water and nutritional deficiencies from spectral signatures. In a second approach, this work studied the capabilities of spectral imaging for crop stress detection. The main objective of this stage was to integrate the spatial information provided by spectral imagery into the framework developed in the first stage. The proposed method was evaluated using images with various spatial and spectral resolutions. The results show that using the full spectral signature instead of vegetation indices significantly improves stress detection. Support vector machines or neural networks using complete spectral signatures obtained detection accura cies of up to 98% for common bean, 88% for maize, and 75% for avocado crops. These percentages vary according to type, stress level, and genotype. The main challenge in using spectral signatures is data collection since it requires extensive fieldwork. As an alternative, we evaluated a methodology with multispectral images of only ten bands, which facilitates data acquisition, achieving 88% and 70% stress detection accuracy in common beans and maize