The ability to detect and map invasive plants to the species level, both at high resolution and over large extents, is essential for their targeted management. Yet development of such remote sensing methodology is challenged by the spectral and structural similarities among many invasive and native plant species. We developed a multi-temporal classification approach that uses unmanned aerial vehicle (UAV) imagery to map two invasive annual grasses to the species level, and to distinguish these from key functional types of native vegetation, based upon differences in plant phenology. For a case study area in the western Great Basin, USA, we intentionally over-sampled with frequent (n=9) UAV flights over the growing season. Using this information we compared the importance of spectral variation at a given point in time (i.e., with and without near-infrared wavelengths), with spectral variation across multiple time periods. We found that differences in species phenology allowed for accurate classification of nine cover types, including the two annual grass species of interest, using just three dates of imagery that captured species-specific differences in the timing of active growth, seed head production, and senescence. Availability of near-infrared imagery proved less important than true-color RGB imagery collected at appropriate time periods. Thus, multi-temporal information provides a substitute for more extensive spectral information obtained from a single point in time. The substitution of temporal for spectral information is particularly well suited to UAV remote sensing, where RGB image collection is inexpensive, and the timing can be flexible. The datasets arising from our multi-temporal classification approach provide high-resolution information for modeling patterns of invasive plant spread, for quantifying plant invasion risk, and for early detection of novel plant invasions when patch sizes are still small. Widespread application and up-scaling of our approach requires 42 advances in our ability to model the variability in phenology that occurs across years and over fine spatial scales, even within a single species.

UAV imagery was collected across the growing seasons of medusahead and cheatgrass between the spring thaw and the summer dry down from May 15th to July 20th, 2017, as frequently as every week (8 total flights – May 15, May 19, June 1, June 7, June 15, June, July 20). The 5-band spectral data were collected with a Micasense RedEdge sensor mounted on a Tarot quadcopter UAV platform flown at an altitude of approximately 30m. In addition to the 5-band Micasense RedEdge sensor we also collected data using a 3-band true color Sony A5100 24.3-megapixel camera with the Blue, Green, and Red bands centered at 475, 551, and 627 nm. Flight times were typically 45 minutes centered around solar noon and were based on pre-programmed flight plans using Universal Ground Control Software (UgCS) version 3.0.1302 software (SPH Engineering, Riga, Latvia) with a single battery swap and a single pass flight pattern. Photographs were captured ensuring 80% frontal overlap and 70% side overlap. The resulting data have 2-cm spatial resolution and spectral resolution of Blue (475 nm center, 20 nm bandwidth), Green (560 nm center, 20 nm bandwidth), Red (668 nm center, 10 nm bandwidth), Red Edge (717 nm center, 10 nm bandwidth), and NIR (840 nm center, 40 nm bandwidth). Eight ground control point tiles were deployed for the duration of the study and used to co-align data from different flight dates and geo-reference the image stack using a Trimble R10 GPS receiver\ with RTX solutions (Trimble Inc., Sunnyvale, California, USA) with horizontal accuracy up to 8 mm and vertical accuracy up to 15 cm.

Composite rasters were processed in Pix4D v. 3.2 (Pix4D, Lausanne, Switzerland) using structure-from-motion (SfM) techniques that mosaicked images across all dates. Although Pix4D created a three-dimensional point cloud that is used to tie images together we did not use the height data in our classification of vegetation. All bands, across all time periods, were co-aligned in Pix4D resulting in an average horizontal positional precision of 1.26 cm. The eight ground control points obtained from the Trimble were used to georeferenced the images into a NAD83 UTM Zone 11 projection. Calibration to reflectance was achieved with an empirical line approach (Smith and Milton, 203 1999) that took advantage of four spectrally diverse field targets that were laid out prior to each flight. Reference reflectance data were collected in the field on July 20, 2017 during cloud-free conditions just prior to the UAV flight using a Spectral Evolution SR-3500 (Spectral Evolution Inc., Haverhill, Massachussetts, USA). The SR-3500 collects data at 3-nm in the visible range. These data were resampled to match the bandwidths of the Micasense RedEdge and the Sony A5100 cameras and a linear regression was used to convert camera digital numbers to reflectance for each band individually. Field vegetation data was collected using sixteen 1-meter square quadrat frames on the day of each flight. To ensure that quadrat frames were visible in each image we permanently placed four metal five-inch spikes at the corner of each frame, and we placed the frames on each of the metal spikes prior to each flight. Quadrat frame locations were initially collected with a Trimble GeoXT GPS unit (Trimble Inc., Sunnyvale, California, USA), and quadrats were re-photographed upon each subsequent visit. Initial vegetation measurements included estimating the aerial cover of each plant species as well as bare soil, rock, animal feces, medusahead litter, cheatgrass litter, and other plant litter. Each quadrat frame was classified into one of the following nine cover types based on its dominant vegetation type: annual forb, cheatgrass, crested wheatgrass, medusahead, litter, perennial forb, perennial grass, shrub, and bare soil. Slight differences in quadrat frame placement were accounted for such that pixels that had quadrat frames visible in them during any of the flights were removed. We randomly selected 9168 points from the resulting polygons to use for model training extracting each spectral band for each flight date. To obtain a fully independent validation we digitized polygons of dominant vegetation types and extracted 258,554 random points to use as validation. We manually digitized polygons using all eight dates of imagery as a background image coupled with extensive field knowledge of the site.

We used the Random Forest algorithm to classify each pixel in the composite image into one of the nine dominant vegetation types. Random Forest is an ensemble decision tree classification in which trees are trained using bootstrapped sampling. Random Forest has been extensively used for image classification because of its high performance, lack of reliance on an underlying data distribution, and its ability to handle both continuous and categorical predictors. For this study we used the RandomForestSRC package in R. Models were run in classification mode using 1000 trees (ntree) and with the number of variables for splitting (mtry) set to the square 11 root of the number of predictors. In order to compare the importance of different spectral wavelengths and different dates on the classification accuracy we ran models for all combinations of dates and bands. Importance values for individual predictors (band*date combinations) were calculated using the mean decrease in impurity (Gini Index). Models were evaluated by creating a classification confusion matrix and assessing overall accuracy, Cohen’s kappa, and by comparing the overall accuracy to that of a null model in which the probability of all classes are equal.

The data are split into three folders. Occurrence contains training and validation points obtained from field sampling that were used to build and validate models derived from UAV imagery. Images contain orthomosaics from each of the eight flight dates and each of the five spectral bands. The classification folder contains the final classified vegetation map using the "best" three dates of imagery.  Each folder contains its own readme file explaining what each individual file is.