A remote sensing-machine learning framework for monitoring vineyard performance
Stellenbosch University, Faculty of Social Science, Department of Geography and Environmental Studies
Specific objectives were to:
1. Develop models for the detection and monitoring of water stress in a Shiraz vineyard using hyperspectral data (HRS) and machine learning (ML).
2. Derive an optimized two-band vegetation index for improved modelling of water stress in a Shiraz vineyard.
The project exploited the utility of narrowband HRS data to improve the detection of water-stressed vines. HRS data was utilised as it has proven to detect disparities more proficiently than multispectral data. Models were also developed using multitemporal data to produce a more robust operational framework. The study employed machine learning approaches and hyperspectral imaging to model water stress in a Shiraz vineyard. Water-stressed vines were discriminated from non-stressed vines by building classification models using leaf spectra samples and tree-based ensemble learners. Accuracy and robustness of analysis were evaluated on independent datasets.
Feature selection methods were employed to identify wavebands most relevant for water stress detection in a Shiraz vineyard. The important wavebands, together with multitemporal data, were then used to develop spectral indices that are optimised to detect water stress in a Shiraz vineyard. The visible (VIS) region of the EM spectrum could successfully be utilised to model water stress in a Shiraz vineyard. Tree-based ensemble learners can successfully identify water-stressed vines in a Shiraz vineyard. Demonstrated the feasibility of terrestrial hyperspectral imagery to model vineyard water stress. Highlighted that optimised narrowband indices outperformed traditional water stress vegetation indices. Identified wavebands that are most relevant for water stress detection in a Shiraz vineyard. This project presents a novel hyperspectral–machine learning framework for the non-destructive identification of water-stressed vines. The results indicated the viability of narrow wavebands to model vineyard water stress and established the utility of tree-based machine learning ensembles within the domain of viticulture. Furthermore, the study demonstrated the feasibility of feature selection methods for the development of optimised spectral indices. The study provides a point of departure for the operationalisation of future machine learning–remote sensing frameworks for water stress monitoring.
A key output of the research is the identification of specific spectral bands that could be used to monitor and evaluate vineyard performance, specifically water stress. These spectral bands can ultimately be used to inform the design of a customised multispectral sensor for rapid, real-time, in-field monitoring of vines. A customised multispectral sensor will be more cost-effective to produce, easier to use, and be employed for a myriad of applications within precision viticulture. The development of a two-band spectral index could serve as a valuable operational tool for in situ water stress detection. The methodology can also be readily applied to a number of other applications in the viticulture sector including yield estimation, stress (pests and disease, drought, etc.) detection, and monitoring nutrient status.
It is recommended that a further study be conducted to evaluate the commercial viability and transferability of the developed methodology. Ideally, the commercial viability study should focus on satellite platforms as it is seen as the most pragmatic solution for commercial application. This can be done by incorporating satellite imagery with similar spectral bands as the ones identified in this study as being most relevant for water stress detection. The commercial viability study should also incorporate multiple cultivars and trellis systems to test the transferability and robustness of the developed methodology. The study can also be replicated by applying the methodology to imagery collected using unmanned aerial vehicles (UAV). Such a study would more than likely utilise multispectral sensors, which is more cost-effective in terms of commercial use. It is worth noting that narrow spectral bands (used in this study) would be the most accurate optical remote sensing solution for the detection of water-stressed vines. However, the instruments (i.e. hyperspectral cameras, spectrometers or spectroradiometers) required to collect such datasets are not viable options for commercial use due to the high cost (often greater than R1 million) associated with these type instruments. Nevertheless, these instruments are ideal for research applications that could potentially produce work that is more cost-effective and commercially viable.
Dukes, M.; Poona, N; Loggenberg, K. Optimised narrowband spectral vegetation index for rapid assessment of vine water stress.
Loggenberg, K.; Strever, A.; Greyling, B.; Poona, N. Modelling Water Stress in a Shiraz Vineyard Using Hyperspectral Imaging and Machine Learning. Remote Sens. 2018, 10, 202. (doi:10.3390/rs10020202)
Loggenberg K & Poona N (in press). A Feature Selection Approach for Terrestrial Hyperspectral Image Analysis. South African Journal of Geomatics.