Individual digitized CIR images provided clues about weed patch distribution that could be used by land managers in preliminary assessments to identify suspect areas to investigate on the ground. In March CIR images, for example, vegetation patches associated with forage dominated ground truth points generally appeared greener than those associated with weed-dominated points . In May CIR imagery, the opposite was true. While less greenness was evident overall, because the landscape was drying, the vegetation patches associated with weed-dominated ground truth were greener than forage-associated patches . However, for stronger identification power, our results demonstrate the value of standardizing imagery to a single spatial scale and to NDVI units, which help correct for differences in illumination. In the northern hemisphere, for example, solar zenith angles are greater in March than in May, and thus produce more notable hill-shadows, which can make interpretation more difficult . These hill shadow effects were essentially eliminated with NDVI —because it is calculated as a ratio of radiation types—resulting in cleaner comparisons of imagery time series and stronger identification of the phenological signatures of vegetation over time. This allowed us to use two images per year with more confidence, either as stacked NDVI or difference NDVI , from which final classifications produced easy-to-read maps .
We analyzed the standardized NDVI imagery to quantify the phenological signatures of our target vegetation types. We extracted NDVI values from imagery at known locations dominated by “pure” forage or weeds . In comparisons of mid-March and mid-May values, hydroponic trays we found significant month x vegetation type effects as predicted , indicating that the annual weed grasses could be distinguished from the annual forage species by the nature of the change in their spectral properties over the growing season. In March , mean NDVI values in weed-dominated patches were less than half those in forage patches, whereas in May the opposite relationship was evident: weed patch values were about 2-fold greater than forage values . However, the highest NDVI values at weed-dominated points only reached 54% of the highest values at forage-dominated points . The NDVI difference imagery provides one simple way to capture these distinct temporal signatures of the vegetation types and showed greater difference between them than either March or May NDVI values alone . The NDVI values at ground-truth points showed additional features consistent with the system’s field ecology, which increased our confidence in the biological value of the digitized imagery. Three examples illustrate this point. First, in the 2009 drought year, NDVI values were reduced at both weed-dominated and forage-dominated ground truth points , with the biggest reductions in NDVI evident in May , which was an effect consistent with our general field observations. Second, we also found that NDVI values at ground truth points varied by property , in a manner consistent with our field impressions.
The lowest mean NDVI values were seen on the least-grazed ranch, particularly in March . This property had extensive patches of weedy grasses with thatch accumulation, so the low NDVI values are consistent with the biological likelihood that thatch build-up obscured the detection of the green biomass in spring, or actually suppressed it, as seen in Bartolome, Stroud. Finally, comparison of the two different dates of image acquisition in May 2009 captured the field-observed quick decline in greenness of the weed-dominated cover in that short period , illustrating the ephemeral nature of that signal.Overall accuracy describes the percentage of all ground-truthed patch types that were correctly identified across the entire mapping area. The kappa statistic is a more rigorous metric, which adjusts overall accuracy for the probability that some of the patches could be correctly identified based on pure chance. Analysis of the distribution of overall accuracy and kappa statistic values for the different mapping approaches indicated that the most important determinant of classification success was the choice of input . For example, we used least squares ANOVA to model the kappa values from all classifications as a function of the fixed effects of year , classification method , and image input type . In this model, image input type had a notably greater effect on kappa statistics than did either year or classification method . Contrast tests among image input types found that classifications derived from time-series inputs containing information from both March and May notably out-performed those based on single images per year . Kappa values for classifications based on time-series inputs exceeded 0.6, which describes good to excellent classifications whereas those based on single images did not .
Results for overall accuracy were similar , with classifications using time-series imagery more readily achieving 80% accuracy . As noted, year effects in this model explained an additional but smaller portion of variability in kappa values. Mean kappa values were marginally greater during 2008 when precipitation was average than in the 2009 drought year . Although classification method itself did not significantly explain additional variability in kappa values , mean kappa values were highest for maximum likelihood supervised classification and lowest for unsupervised classification . In a second similar analysis, we considered how kappa values for 2009 classifications were influenced by the choice of May image acquisition date, using a modified model with fixed effects of May date , classification method, and image input type . In this model, the effects of classification method and image input type were both significant . More notably, classifications based on the later May date had significantly poorer kappa values . Moreover, the lowest kappa value and the only statistically unsupported classification were based on May B imagery. In our analyses, we sought to identify a mapping approach that would both achieve high mean kappa and perform consistently well across time. Thus, for each approach we also evaluated the coefficient of variation across time as a function of mean kappa, and judged the best performers to be those with both high mean kappa values and low kappa CV . This evaluation underscores the value of classification based on multiple seasonal NDVI images , as opposed to one-time shots. Contrary to initial field based intuition, classifications based solely on May images were the worst performers . By our criteria, the best performer was maximum likelihood supervised classification of stacked NDVI . Maps produced by this approach showed fine details of weed- and forage-dominated areas and corresponded well to ground truth . Even the “runner-up” methods produced valuable classifications, particularly when based on two images. The degree of detail captured by these maps was notable. In additional informal tests, we loaded maps onto our GPS units and were able to walk through the grasslands while watching in real-time our progress across the digital maps. One of the features that these maps captured was the persistence of large nearly monospecific weed-dominated patches in the ungrazed management units, which we also observed in the field but could not quantify over a large area without the assistance of remote sensing methods. As illustrated , some of these patches were tens, sometimes almost hundreds, of meters across, and varied little across the two years, likely because of the persistence of accumulated thatch and thatch feedbacks. In this ungrazed, weed-dominated landscape, distinct forage patches were also evident but during the drier year these patches lost some integrity and were invaded by the weeds . Comparison of May imagery dates shows that choice of date might influence how the borders and homogeneity of patches were perceived, but general patch patterns remained recognizably similar.Since our primary aim was to map weed cover, it was vital to assess the probability of locating all of the existing weed-dominated vegetation known from ground-truth when using a given map . The percentage of weed-dominated cover a map misses can then be estimated by subtracting the producer’s accuracy from 100%. As with the kappa statistic, pipp mobile systems we considered both the mean and the CV of the producer’s accuracies for the different mapping approaches over time. In this evaluation, the maximum likelihood supervised classification conducted on stacked NDVI was again the strongest performer, with mean producer’s accuracy of 93% and CV of 2.2% , indicating this approach was likely to miss only 7% of weed-dominated cover. Parallelpiped supervised classifications of stacked NDVI and ΔNDVI were consistent performers as well.
The capacity to map all existing weed-dominated cover must be balanced with the chance of misidentifying other vegetation as weed-dominated, so we also considered the user’s accuracy for weed-dominated cover. User’s accuracy estimates the purity of on-the-ground vegetation within the mapped class, and the percentage of other vegetation misidentified as weed-dominated can be determined by subtracting the user’s statistic from 100%. Mean user’s accuracy for all classifications was good with low CV indicating that the maps generally did not misidentify other vegetation as weed-dominated . For the maximum likelihood supervised classification—the best performer in terms of kappa and producer’s accuracy—the mean user’s accuracy was strong as well. The strongest performer for this metric was unsupervised classification of difference imagery, with a user’s accuracy of 100%.Once we had identified the most accurate and consistent mapping approach , we applied it to our study site landscape, which included four management units from three different independent ranches. We evaluated the distribution of weed-dominated cover in both 2008 and 2009, and the nature of any changes between years. In this analysis, 11.6% of the 6.8-km2 study landscape contained non-target land cover types such as trees, row crops, or non-vegetated surfaces , which we masked in the imagery and did not examine. The remaining landscape area represented grasslands, which was categorized as weed-dominated, forage-dominated, or neither . We then combined the vegetation maps from 2008 and 2009 to evaluate the nature of vegetation change or stability . The geopositional correspondence of the annual maps was excellent, due to the effective positional accuracy of the ground control points , the fine-grain nature of the original imagery, use of a fine-scale digital elevation model in image orthorectification, and uniform upscaling to 1-m pixels for NDVI image production. Together, the fine-scale individual year maps and vegetation change map provide opportunities for numerous rich analyses. Overall, assessment of 2008–2009 vegetation change showed that 56% of the grassland area across the study site remained dominated by the same vegetation type in both years, while 28.3% converted from one type to another, and 15.4% was unclassified in at least one year . To demonstrate applications, we highlight some aspects of how weed cover dynamics varied among management units and in response to interannual environmental variability. Most notably, the vegetation change map and summary values indicate that the proportion of weed-dominated cover in both 2008 and 2009 varied significantly among management units . In Figs 7 and 8, cover that was weed-dominated in 2008 is the sum of the red + blue areas; cover that was weed-dominated in 2009 is the sum of the red + yellow areas. Specifically, in both years weed cover dominated proportionally more of the grasslands in management units with no significant grazing than in units with repeated grazing . When grazing was absent, large near-monotypic weed patches dominated, often with organic shapes that followed landscape contours. In the field, these weed patches were associated with accumulated layers of thatch. The influence of management on weed accumulation is highlighted by fence-line contrasts, such as the border between M3 and M4 , for which soil types are identical on each side. Close analysis of patch details found that many weed patches spread outwards in bands at their margins during the 2009 drought year , while smaller and more scattered forage patches sometimes emerged within them. In contrast, weed patches in grazed units were more varied in shape and location.A key goal in ecological research is to understand what factors control the spatial distribution of vegetation patches and whether patches persist, expand, or contract over time. Despite technological development, data to assess fine-grained patterns of vegetation turnover at the landscape scale remain limited. This information gap hampers efforts to establish and prioritize management options for invasive species control. Landscape-scale information about patch dynamics within grassland communities is particularly lacking, because many grassland shifts occur between species of the same functional type and have been difficult to detect remotely .