Scalable asymptomatic grapevine leafroll virus complex-3 detection through integrated airborne imaging spectroscopy, autonomous robotics, and cloud computing

Context & Purpose

The past three decades of terrestrial remote sensing research have delivered unprecedented insights into our fundamental ability to detect, quantify, and differentiate plant disease (Gold 2021). However, much of our fundamental knowledge in this domain has come from studies in non-agricultural systems and until recently, most agricultural studies, when extant, have focused on tree crops where canopy closure and large plot and plant size facilitate stress detection at low spatial resolution. Recent engineering innovations and advancements in constellation architecture design have refined the accuracy and scalability of airborne and spaceborne sensing platforms, enabling us to monitor diverse specialty crops, including grapevine, planted in smaller, spatially varied fields. Most prior work on grapevine remote sensing has aimed to detect water stress, with few studies addressing other abiotic parameters such as nitrogen content, yield, and fruit composition. Reports on proximal and remote sensing of grape disease have been focused on visible disease detection and predominantly limited to near surface deployments (Naidu et al. 2009; Oerke et al. 2016; Hou et al. 2016; MacDonald et al. 2016; Knauer et al. 2017; Bierman et al. 2019; Bendel et al. 2020; Gao et al. 2020; Lacotte et al. 2022; Sawyer et al. 2023). Plant disease alters how solar radiation interacts with leaves, canopy, and plant energy balance resulting in changes that are readily capturable in visible to shortwave infrared (VSWIR, 400-2400nm) hyperspectral imagery prior to symptom appearance (Gold et al. 2020). Exciting recent work has established that airborne imaging spectroscopy is capable of pre-symptomatic disease detection in multiple culturally important pathosystems, including olive quick decline syndrome (Zarco-Tejada et al. 2018, 2021) and infections of oak by Phytophthora spp. (Hornero et al. 2021) and Bretziella fagacearum (Sapes et al. 2022). This novel capacity combined with the anticipated spectral, spatial, and temporal coverage of forthcoming hyperspectral satellite systems, such as NASA’s Surface Biology and Geology (SBG, Schneider et al. 2019) and ESA’s Copernicus Hyperspectral Imaging Mission for the Environment (CHIME, Nieke and Rast 2018), will revolutionize imaging spectroscopy data availability for agricultural decision making, enabling disease monitoring at hereto unachievable resolutions while providing a robust foundation for multiscale risk estimation and surveillance networks.

While these studies in aggregate have proved imaging spectroscopy to be powerful tool for understanding, detecting, and mapping plant-fungal, oomycete, and bacterial interactions at scale, viral-plant interactions have yet to be explored at suborbital and beyond scales. Viral diseases, including that caused by Grapevine Leafroll Virus Complex 3 (GLRaV-3), cause $3 billion in losses to the US wine and grape industry annually (Naidu et al. 2014). GLRaV-3 infection significantly reduces vine longevity and causes the grapevine to misappropriate resources, resulting in uneven cluster ripening, changes to grape berry chemistry, reduced wine quality, and visible foliar reddening post-verasion in red grape varieties. Further challenging management is GLRaV-3’s long (~12-month) latent phase, during which the host is infectious, but foliar symptoms are not yet apparent (Naidu et al. 2014; Blaisdell et al. 2016). Existing strategies to detect GLRaV-3 in the field are predominantly based on visual scouting by trained experts, a labor and time intensive process. This means that both latently infected red grape varieties and white grape varieties (which do not manifest visible foliar symptoms) serve as inoculum sources for nearby vineyards without grower recourse other than expensive ($40-50 per vine) molecular testing. Previous work has demonstrated the utility of remote sensing data for large-scale detection of symptomatic GLRaV-3 in red grape varieties (Hou et al. 2016; MacDonald et al. 2016)however these proof of concept studies have not yet transitioned from research to application, nor have they explored asymptomatic detection. Effective asymptomatic detection is critical for management, especially when the practicalities of deploying scouts across the hundreds of thousands of acres of grapevine grown in the United States is considered. Even under the most aggressive scouting plans, it takes at minimum 12 months, and more frequently, 18-24+ months, from date of first infection to date of removal.

Vineyards are difficult to monitor via remote sensing because of limited equipment viewing angle that challenges side and lower canopy symptom detection. The grapevine canopy is small relative to the image footprint of most air- or spaceborne sensors, and vines are bordered by inter rows of bare soil and/or non-vine vegetation, and shadows cast by the crop itself impart noise to the reflectance signal. Effective remote sensing requires large amounts of calibration and validation data, which can be challenging to obtain through traditional human field scouting. This has led to widespread interest in automated ground-based detection methods for early detection, and various techniques for data processing ranging from traditional statistical methods to deep learning models developed (Naidu et al. 2009; Bendel et al. 2020; Gao et al. 2020; Sawyer et al. 2023). Thus far, use of these methods has been limited to situations where leaves or branches are sampled and placed against a controlled background. The logical next step to bridge the research to application gap is deployment on uncrewed ground vehicles (UGV) capable of traversing in between rows without human intervention to collect validation data for aerial imagery at scale. However, vineyards in different geological locations have varying environmental challenges and management requirements. Designing robotic systems suitable for different environments, terrains, slopes, and tasks is critical for simplified, effective deployment and cost control.

Effective use of the afore described technologies is facilitated by machine and deep learning, which help to make sense of the underlying relationships within and amongst hundreds of highly correlated spectral bands. These approaches have been widely used in airborne plant-microbe interaction sensing as well as in the broader, adjacent domains of foliar functional ecology (Townsend et al. 2003; Sousa et al. 2021; Wang et al. 2020). Together these can support agricultural decision-making at unprecedented scales, but training and deploying these models is limited by the considerable compute and storage resources they require, severely limiting non-expert use. Cloud computing offers on-demand and nearly unlimited access to storage and computing resources—where connectivity allows. In agricultural areas with restricted internet connectivity, edge computing complements cloud computing by placing storage and computing nearer to data sources, enabling accurate models to be built and timely response. This distributed computing model enables models to be trained in the cloud and deployed at the edge (e.g.  on users’ devices), optimizing utility for agricultural decision-making (Rubambiza et al. 2022). These methods have been increasingly applied in agricultural remote sensing, including yield prediction, soil mapping, and land management (Rubambiza et. al 2023). Thus, edge and cloud computing are crucial for scaling current techniques and investigating the implications of future spaceborne missions for effective disease surveillance.

NASA’s Airborne Visible/Infrared Imaging Spectrometer Next Generation (AVIRIS-NG) is an airborne instrument operated from the Jet Propulsion Laboratory in Pasadena, CA with extensive historic acquisitions in California, including over ~364,000 hectares (ha) of vineyards. The AVIRIS mission family, which includes Classic (C), Next Generation (NG), and the forthcoming AVIRIS-3, will continue to collect wide-swath, high spectral resolution (<10 nm), and highly uniform spectroscopic imagery (400 to 2400 nm) over diverse California biomes, including viticultural production areas. Additionally, NASA’s forthcoming hyperspectral satellite Surface Biology and Geology will be based on the AVIRIS family architecture. AVIRIS-NG therefore represent perfect opportunity to generalize, optimize, and continuously validate models for early GLRaV-3 detection in vineyards. Thus, the goals of the work presented here are threefold:

• Determine to what level of accuracy airborne imaging spectroscopy can detect asymptomatic GLRaV-3 infection across scales.
• Deploy these models in a distributed, cloud computing system designed with stakeholder values and needs in mind.
• Deploy an autonomous ground robotic training system to collect ground truth validation at scale in real time to improve model development.