UAV imagery was collected in April and July of 2017 over a grape vineyard in California’s San Joaquin Valley. Using spectral signatures, a landcover classification was performed to isolate table grapes from the background vegetation and soil. A novel vegetation index was developed based off the unique spectral characteristics of the yellowing effects of chlorosis within the table grape vines. Spatial statistics were run only on the pixels containing grape plants, and a relative vegetation health map was produced, using both the novel vegetation index and NDVI. Several regions within the map were identified as areas of interest and marked for further analysis. Automated temporal analysis showed that these areas of interest were in fact underperforming throughout the growing season. Field measurements later confirmed that these areas were suffering from a potassium deficiency. It was also shown that using a simple vegetation index, like NDVI, alone could not locate these anomalous areas even though it was present and measurable within the data.

In 2017 canola was planted on 9 million hectares in Canada surpassing wheat as the most widely planted crop in Canada.  Saskatchewan is the dominant producer with nearly 5 million hectares planted in 2017.  This crop, seen both as one of the highest-yielding and most profitable, is also one of most expensive and input-intensive for producers on the Canadian Prairies.   In this study, the effect of natural and planted shelterbelts on canola yield was compared with canola yield from fields with no tree or other natural vegetation. Yield was measured in 15 canola fields, 5 of which had naturally occurring trees, 5 with a row of planted trees and 5 without any trees, along three, 350 m transects, near Indian Head, Saskatchewan.  During the growing season RGB and radiometrically corrected multispectral data from an unmanned aerial vehicle were collected at least once at each site.  A variety of vegetation indices, such as normalized difference vegetation index, enhanced vegetation index, soil adjusted vegetation index and other variables were correlated with ground based measurements to determine the impact tree rows have on yield.  Preliminary results from one replicate site which includes the natural shelterbelt, planted shelterbelt and control site are described here.  A weak and inconstant correlation was found between yield and distance from natural vegetation.  

This paper presents a method which can provide support in determining the grass-clover ratio, in grass-clover fields, based on images from an unmanned aerial vehicle. Automated estimation of the grass-clover ratio can serve as a tool for optimizing fertilization of grass-clover fields. A higher clover content gives a higher performance of the cows, when the harvested material is used for fodder, and thereby this has a direct impact on the dairy industry. An android application is implemented to make the drone fly fully autonomously and collect images at different locations within the field. In this android application it is possible to specify what location the drone should collect images from, which height, and upload the images to a server, which analyze the data based on a convolutional neural network. The convolutional neural network performs a semantic segmentation and thereby pixelwise classify the different classes: grass, clover, soil and weed. The classification of the pixels is used to determine the final grass-clover ratio. The results, presented in this paper, show that the CNN is able to segment the images into the different classes: grass, clover, soil and weed. It is possible to identify the different classes based on images captured at a height up to five meters. Thus, this paper shows a way to use UAVs to perform mapping of actual clover and grass ratio in dense grass-clover fields.

Previous work (Binch & Fox, 2017) used autonomous ground robotic platforms to successfully detect Urtica (nettle) and Rumex (dock) weeds in grassland, to improve farm productivity and the environment through precision herbicide spraying. It assumed that ground robots swathe entire fields to both detect and spray weeds, but this is a slow process as the slow ground platform must drive over every square meter of the field even where there are no weeds. The present study examines a complimentary approach, using unmanned aerial vehicles (UAVs) to perform faster detections, in order to inform slower ground robots of weed location and direct them to spray them from the ground. In a controlled study, it finds that the existing state-of-the-art (Binch & Fox, 2017) ground detection algorithm based on local binary patterns and support vector machines is easily re-usable from a UAV with 4K camera despite large differences in camera type, distance, perspective and motion, without retraining. The algorithm achieves 83-95% accuracy on ground platform data with 1-3 independent views, and improves to 90% from single views on aerial data. However this is only attainable at low altitudes up to 8 feet, speeds below 0.3m/s, and a vertical view angle, suggesting that autonomous or manual UAV swathing is required to cover fields, rather than use of a single high-altitude photograph. This demonstrates for the first time that combined aerial detection with ground spraying system is feasible for Rumex and Urtica in grassland, using UAVs to replace the swathing and detection of weeds then dispatching ground platforms to spray them at the detection sites (as spraying by UAV is illegal in EU countries). This reduces total time requires to spray as the UAV performs the survey stage faster than a ground platform.

Cotton is an indeterminate crop; therefore, fertility management has a major impact on the growth pattern and subsequent yield. Remote sensing has become a promising method of assessing in-season cotton N status in recent years with the adoption of reliable low-cost unmanned aerial vehicles (UAVs), high-resolution sensors and availability of advanced image processing software into the precision agriculture field. This study was conducted on a UGA Tifton campus farm located in Tifton, GA. The main goal of this study was to correlate in-season cotton N status with multispectral imagery acquired with a UAV. For this study, six N treatments consisting of 0, 34, 67, 101, 135 and 168 kg/ha rates were applied to attain varying levels of plant N status within the same field. Cotton tissue samples were collected during crop growth stages (first, third, fifth, and seventh week of bloom) to quantify plant N status during these stages. Tissue analysis results provided leaf blade N (%) and petiole N (ppm) for each N treatment implemented in the study. Crop multispectral imagery in the spectral wavelengths of 550 nm (green), 660 nm (red), 735 nm (red-edge) and 790 nm (near infrared) was acquired during these crop growth stages by utilizing a commercially available quadcopter equipped with a high-resolution multispectral camera. Different vegetation indices (VIs) were selected and calculated based on potential correlation with plant N status and were calculated from the data acquired from the multispectral aerial imagery. Correlations between the indices and leaf blade N (%) and petiole N (ppm) as obtained from plant tissue analysis were compared. Regression equations correlating the VIs to actual N levels were generated to evaluate the use of different VIs for accurately measuring N levels in the crop at the selected growth stages. Initial data analysis indicated that NDVI was strongly correlated to leaf blade N (%) and petiole (ppm) from the first week of bloom samples, whereas, NDRE had stronger correlation for the samples that were taken in the third, fifth, and seventh week of bloom. These correlations may provide promise for using multispectral imaging to detect in-season N variability in cotton.  

Intra-field variability has received much attention in arable and horticultural contexts. It has resulted in increased profitability as well as reduced environmental footprint. However, in a pastoral context, the value of understanding intra-field variability has not been widely appreciated. In this programme, we used available technologies to develop multiple data layers on multiple fields within a dairy farm. This farm was selected as it was already performing at a high level, with well-developed existing infrastructure and high production metrics. All of the fields on this farm were around 4 hectares and had established pastures based on perennial ryegrass/white clover. We developed high resolution data layers on pasture height, soil electroconductivity and soil texture at a pixel size of approximately 4m. Maps of the pasture height of individual fields sampled at the same interval post-grazing (~21 days) were then combined to produce a whole-farm map of relative pasture production. This revealed a strong interaction between irrigation performance, soil bulk density and pasture growth. These patterns had not previously been noted by farm management. In addition, several zones were identified that showed lower relative pasture yield, subsequently demonstrated to be due to high densities of soil-dwelling pasture pests (scarab larvae). As a result of this information, management decisions were made to improve irrigator performance – especially the uniformity of water application – and to address the pest problem by applying a prototype biopesticide. Modelling of the improvements in pasture production that were expected as a result of these decisions predicted an increase in total pasture production to more than 20 t DM/ha/yr (~10%). This exceeds the accepted pasture production expectations for this region and sets a new target for irrigated dairy farms.

Tree geometry such as volume, height, and width are important information that can help growers to conduct a precise orchard management. Conventionally, canopy profile is generated by using light detection and ranging (LIDAR) technology as a method for direct measurement of tree structure. While LIDAR is a precise method for generating 3D models of trees, it requires expertise, and expensive equipment that limits its application for creating 3D maps in large orchards. In this paper, an affordable and simple method for generating a 3D reconstruction of an orchard, called Virtual Orchard, was introduced and its capabilities were evaluated. Virtual Orchards for almond, citrus, pistachio, and vineyard were created by photogrammetry technique using a series of aerial images of orchard/vineyard acquired by an unmanned aerial system from various angles and heights. Orchard topography and digital surface model (DSM) were extracted from a point cloud dataset created by Virtual Orchard technology. An algorithm was developed to identify and locate individual trees in a virtual orchard and extract information such as canopy cover, maximum height, average height, area, width, and volume index. This information can be further analyzed and interpreted by prediction models to generate useful and timely prescription maps for precision agriculture practice. Virtual Orchards were also developed in non-visible bands such as near infrared (NIR, at about 800 nm) and red edge (680-740 nm). NIR imagery indicates plant vigor, in which a lower NIR value usually means that plant is under stress. Additionally, NIR could be used for normalizing chlorophyll-sensitive bands (e.g. Red band) and creating vegetation indices such as NDVI. Red-edge is another non-visible band that includes important information about plant stress, chlorophyll, and nitrogen content. Lower values in red-edge band usually indicate that plant is under nitrogen stress. Virtual Orchard in non-visible band could illustrate plant reflectance not only on tree crowns, but also on the sides of each tree. The virtual orchard technology demonstrated the potential to help growers conduct a precise management to improve yield, decrease waste, and maintain the quality of the environment.