Leafy spurge (Euphorbia esula L.) is a perennial invasive weedy species that has successfully established across a wide range of ecosystems and has a noxious weed designation in 22 US states (Goodwin et al., 2001). Leafy spurge is difficult to eradicate, despite heavy use of herbicides, targeted grazing, and various biocontrol agents (Goodwin et al., 2001). Leafy spurge is adapted to a wide range of growing conditions, including disturbed areas, rangelands, pastures, and river bottoms (Hyder et al., 2008).

With blue-green stems, linear leaves, small green flowers, and bright yellow bracts, mature leafy spurge is easily identifiable. Individual stems can produce over 200 seeds annually, which when mature, can be expelled up to 15 feet from the plant (St. John and Tilley, 2014). Seeds are small and buoyant, easily transported by water to areas downstream. Tissues contain a milky white sap that contains ingenol, which is released when plants are injured (Larry Leistritz, 2004). Once ingested, leafy spurge can be toxic to livestock, horses and wildlife, and result in significant economic losses to producers and other land users (Goodwin et al., 2001).

Remote sensing for leafy spurge mapping has been conducted since 1995, when leafy spurge was mapped with real color aerial imagery at Theodore Roosevelt National Park and found to be associated with high proximity to waterways (r ² = .98) (Anderson et al., 1996). Multispectral sensors, which record reflectance beyond the visible spectrum in 3–10 spectral bands, have been used in a mixed-grass prairie in North Dakota (Casady et al., 2006) Leafy spurge was mapped with 67% accuracy, with decreasing classification accuracy when grown with herbaceous vegetation (Casady et al., 2006).

Successful remote detection of leafy spurge depends on capturing spectral differences between leafy spurge and other land cover, and imagery with limited spectral bands can make identifying leafy spurge challenging (Hunt andParker Williams, 2006). In a heterogeneous landscape near Devil’s Tower National Monument Crook County, Wyoming, hyperspectral (remote sensing imagery consisting of hundreds to thousands of spectral wavelengths) classifications performed best in prairie vegetation, and better in bottomlands than uplands, which was caused by dense leafy spurge cover in riparian areas and longer periods of blooming in moist habitats (Williams and Raymond, 2002). In a study by Parker and Hunt. (2004), leafy spurge was well correlated with higher near infrared (NIR) reflectance, which is also known to be well correlated with dense vegetation cover. Leafy spurge was detected with 95% accuracy, and classification performed well in mixed-prairie and riparian vegetation than in forested areas (Williams and Raymond, 2002). In the sagebrush (Artemisia tridentata subsp. tridentata) steppe of Swan Valley, Idaho with 84%–94% accuracy (Glenn et al., 2005) and with 96%–99.5% accuracy at the mixed-grass prairie at Theodore Roosevelt National Park (O’Neill et al., 2000).

Leafy spurge has been spreading along the Yampa River main channel and constructed irrigation ditches for over 40 years (Yampa River Leafy Spurge Association, 2018; Turnage, 2021). The Yampa River flows for 250 river miles in Moffat and Routt Counties, Colorado, United States. Increased presence of leafy spurge in this area can be indicative of disturbance and large flooding events (Yampa River Leafy Spurge Project, 2018; Goodwin et al., 2001). The Yampa is one of the last free-flowing rivers in the Western United States, with seasonal flooding temporarily inundating sandbars and terraces away from the active river channel (Merritt and Cooper, 2000). Understanding the current extent of leafy spurge invasion with high spatial resolution is critical as its presence has already been observed beyond the riverbanks and floodplains, such as upland meadows, hillslopes and mountain ridges. High spatial resolution is a priority, as small satellite clusters of leafy spurge away from the main invasion are more likely to be effectively managed and to spread to new areas (Westbrooks, 2004). This project aimed to: 1) produce detailed maps of the current leafy spurge infestation along the Yampa River corridor with SPOT seven satellite remote sensing and 2), visit locations and validate leafy spurge presence or absence to describe classification performance.

The classification method identified mapped leafy spurge training samples with 95.4% accuracy training and an overall classification accuracy of 90.7%. If mapping was to take place again however, satellite imagery with additional wavelengths of near infrared spectra may be useful, as SPOT seven imagery NIR only covers the 760–890 nm of the spectrum. Near infrared reflectance was the only spectral band that was significantly different between correctly identified and missed leafy spurge polygons, so more bands of varying wavelengths of near infrared may be useful in distinguishing leafy spurge’s unique spectral signature. Leafy spurge does seem to be spectrally identifiable within the narrow portion of the electromagnetic spectrum captured by multispectral satellite imagery products (Mitchell and Glenn, 2009). Our detection rates of leafy spurge varied, as infestations that consist of a single leafy spurge plant, leafy spurge growing with trace amounts of bare ground, leafy spurge growing on banks, and leafy spurge growing with riparian shrub vegetation cover types were more likely to be missed by our Random Forest imagery classification. This may be because the presence of a single leafy spurge population is much harder to detect with 1) somewhat limited spatial resolution (4 m × 4 m pixels), 2) changes to bank geomorphology caused by seasonal flooding during the period between satellite imagery collection and validation mapping, and 3), dense shrub cover obscuring leafy spurge invasions beneath their canopy. Overstory canopy coverage alone, however, was not a significant predictor of leafy spurge classification accuracy.

In other research projects where remote sensing was used to map leafy spurge, leafy spurge growing below overstory canopy coverage, leafy spurge on steep slopes, sparse leafy spurge populations, and small infestations of leafy spurge were missed (Anderson et al., 1996; Parker and Hunt, 2004; Glenn et al., 2005), aligning with our decreased detection accuracy of singly leafy spurge plants and leafy spurge growing below shrub and tree canopy. In contrast to our challenges of mapping leafy spurge in riparian areas of the Yampa River, detection accuracy of leafy spurge mapping in Crook County, Wyoming was improved when growing in riparian areas (Williams and Raymond, 2002; Parker and Hunt, 2004). Adding ancillary spatial data, like vegetation cover maps and distance to waterways may improve remote sensing classification results (Hunt et al., 2010; Dubovik et al., 2021). Additionally, the incorporation of citizen science data of dense, unobscured leafy spurge presence collected and submitted from within the study area could be incorporated, to increase training samples and obtain samples away from the Yampa River riparian zone (Vaz et al., 2019) but increasing the temporal resolution of the imagery with a times series of multiple imagery scenes through the leafy spurge bloom could.

One major limitation of this project is the difference between June 2021 validation mapping and satellite imagery data collection in July of 2019. Imagery procurement was limited by YRLSP budget, and 2021 SPOT seven satellite imagery purchase and classification would have been beneficial, as leafy spurge invasions may have advanced. If further remote sensing was to be conducted for mapping general leafy spurge invasion in the area, free coarser resolution multispectral imagery like Sentinel-2 (10 m × 10 m pixels) could be used. As leafy spurge is a deep-rooted perennial and a prolific seed producer (St. John and Tilley, 2014) and largely uncontrolled in the Yampa River Valley, leafy spurge populations away from the dynamic riverbanks are unlikely to decrease, excepting extreme environmental conditions, like drought. In experimental plots where herbicide was applied in the fall of 2019, treatments that controlled leafy spurge reproduction left woody stems which were recognizable during validation mapping in the summer of 2021, so some signs of past leafy spurge may be recognizable even if plants do not regrow from established roots. Additionally, our training samples were biased toward dense, large, and unobscured leafy spurge recognizable in multispectral imagery sets and may not represent the full range of leafy spurge infestation abundance, size, and habitat type.

Though the success rates of the validation mapping efforts were generally low (59% of validation locations were correctly classified), the validation locations were chosen based on anomalies in YRLSP members expert knowledge of the area or features and areas of interest from the classification prediction. For example, a series of validation points were set in a seasonal Yampa River tributary, to see if positively classified pixels as leafy spurge were true leafy spurge locations. Leafy spurge detection accuracy varies within the 190 leafy spurge positive presence locations within the dataset, as the Random Forest classification was more accurate at identifying leafy spurge populations growing as discrete patches rather than scattered populations. Discrete patches of leafy spurge and infestations with higher leafy spurge percent cover were more likely to be correctly classified by the Random Forest imagery classification. Casady et al. (2006) found that detected leafy spurge patches had higher average leafy spurge percent cover and higher average patch size, with infestations smaller than 200 m² and 30% leafy spurge cover less likely to be identified (Casady et al., 2006). Dense populations with high leafy spurge cover may have more recognizable spectral signatures than sparse populations. Discrete boundaries of leafy spurge patches may be more identifiable, as scattered populations might share pixel space with other land cover types, though the number of additional species present at validation locations did not significantly influence classification accuracy of leafy spurge.

Satellite remote sensing appears to be a viable option for leafy spurge mapping on the Yampa River, but classification accuracy varies with leafy spurge infestation characteristics and environment. Small and/or sparse infestations of leafy spurge may go undetected, and detection rates decreased in areas with shrubs and overstory canopy coverage. Identification was most accurate when leafy spurge coverage was dense. Validation mapping suggests that leafy spurge is best identified when growing in discrete patches and when leafy spurge densities are higher.

If leafy spurge mapping was to be conducted again in the Yampa, a time series of imagery that represents early season and mid-season would be employed to try to differentiate leafy spurge from other vegetation, as suggested by Lake et al. in a heterogeneous Minnesota landscape (Lake et al., 2022). If left unmanaged, leafy spurge may become a much more prevalent invader in Moffat and Routt Counties, producing more propagules to spread through the Yampa, Green, and Colorado Rivers.