There is a large and increasing availability of vegetation plot data, particularly in Europe, covering wide spatial extents as well as long-term temporal gradients, largely resulting from decades of field surveys conducted by hundreds of surveyors. These data are fundamental to understand vegetation patterns and the drivers shaping them, including global changes (Chytrý et al. 2016; Bruelheide et al. 2019; Knollová et al. 2024). However, even when backed by botanical and field expertise, sampling relies on visual and inherently subjective judgements made by human observers, both in terms of species identification and estimation of relative abundances (Vittoz and Guisan 2007; Morrison 2016; Salvatori 2024). Owing to this subjectivity, vegetation surveys are prone to several forms of observer-related error, which can substantially affect the reliability and comparability of ecological data (Boch et al. 2022). For instance, following Kéry and Gregg (2003), detectability is defined as the probability of recording a species given its true presence in the plot. Imperfect detectability can generate false absences, leading to spurious estimates of local extinctions and vegetation change if the analyses do not explicitly account for this bias. Futschik et al. (2020) found that observer-related errors can affect our interpretation of the effect of climate change on alpine grassland, specifically, they found differences of 5 to 13% of species in the records conducted by different observers. Similarly, Verheyen et al. (2018) showed that a substantial proportion of the floristic changes attributed to eutrophication or climate are sensitive to detection errors in the herbaceous layer. Observer error is therefore a pervasive issue in vegetation science, particularly in long-term monitoring programmes and multi-observer surveys (Dengler et al. 2011; Bruelheide et al. 2019; Seidling et al. 2020; Morrison et al. 2024), affecting estimates of biodiversity, turnover, and trends. Although it is evident that observer error in vegetation surveys can influence the interpretation of our results, there is still limited awareness of this issue, and the strategies to reduce or control this source of error are rarely implemented (Kercher et al. 2003; Morrison 2021).

One way to quantify observer error is to compare observations with the true value of the variable of interest. However, in vegetation studies, true values are rarely known because a complete and exhaustive list of species in a given area is often unavailable (Di Biase et al. 2025). While accuracy—the closeness of an observation to the true value—is therefore difficult to assess (Salvatori 2024), precision—the degree to which repeated observations agree with each other—can be explicitly evaluated and provides a meaningful measure of observer performance. Observer-related sampling bias can be investigated by examining variation in species richness, species composition and species abundances. Morrison (2016) identified three main types of observer error in vegetation surveys: (i) overlooking error, when species present in the plot are not detected; (ii) misidentification error, when species are incorrectly identified; and (iii) visual estimation error, when species abundances (e.g., in terms of projected ground cover) are inaccurately assessed. Other sources of error are related to “attributes associated with observers” and could include errors (typos) made during data digitalisation due to mental or physical fatigue. When repeated surveys by the same observer yield different results, this is referred to as intra-observer error; when discrepancies arise among different observers or teams, it is referred to as inter-observer error. Focusing on species as observational units, several studies have shown that overlooking small, rare or inconspicuous species inflates apparent turnover among observers (Nilsson and Nilsson 1985; Morrison 2016). This issue, also called pseudoturnover, has been defined as the proportion of species either missed or misidentified by an observer or team between two sampling occasions or among different observers surveying the same plot. Vegetation surveys often include species abundances attributed to different vertical layers, e.g., in forest ecosystems. Vertical stratification is generally recorded more consistently for dominant tree layers, whereas shrub and herbaceous layers are more prone to both intra- and inter-observer variability (Kennedy and Addison 1987; Klimeš et al. 2001; Kéry and Gregg 2003; Archaux 2009). High variability among observers in estimating percentage cover has been widely documented, with systematic tendencies to under- or overestimate cover depending on size of individuals and plot density (Morrison et al. 2020).

In this context, Boch et al. (2022) reported that “it remains unclear whether observer-related pseudoturnover declines with time within a group of observers, thanks to their continuous training in species identification and their increasing experience in conducting vegetation records”. Morrison (2016) and Morrison et al. (2024) confirmed that the use of multiple observers, additional training including active feedback approaches, and continual evaluation and calibration among observers are recommended strategies to reduce observer error in vegetation surveys.

Forest ecosystems represent one of the most frequently studied habitats in observer error research (Morrison 2016), highlighting the relevance of this issue for forest monitoring programmes. Improving our understanding of observer-related precision is particularly important for long-term monitoring schemes involving numerous observers and repeated surveys over time, such as the ICP Forests Level II network (Allegrini et al. 2009; Ferretti 2021) and the Italian National Forest Inventories (Gasparini et al. 2009; D’Amico et al. 2025).

Using data collected during training and intercalibration sessions carried out within these programmes in Italy, we aim to assess whether targeted training and collective briefing can reduce observer-induced pseudoturnover in forest vegetation surveys. Specifically, we estimate whether inter-observer variability decreases between pre- and post-training exercises in terms of species richness and species composition, the latter measured by three dissimilarity indices.

Species richness showed no consistent change between EX1 and EX2 in 2023 (0.92 species, 95% posterior quantile interval: -4.42–5.79) and 2025 (-1.65 species, 95% posterior quantile interval: -2.87 – -0.41), while the precision of species richness estimates increased across observers in both years. Inter-observer variability was higher before training, as indicated by the posterior distribution of the EX2 - EX1 contrast (Figure 2), with a probability of a decrease in inter-observer variability of 70% in 2025 and 67% in 2023.

Dissimilarity analyses revealed a general reduction in inter-observer dissimilarity after training when using Jaccard index and Euclidean distance in both years. In contrast, Bray–Curtis dissimilarity decreased only in 2023, whereas there is evidence of an increase in 2025 (Figure 3). The percentage reduction in dissimilarity ranged between 10 and 30%, with the exception of Bray–Curtis in 2025 (Figure 4).

Training and intercalibration substantially reduced inter-observer variability in species richness and dissimilarity in the 2023 and 2025 exercises. In our case study, the briefing session seems to have produced a similar effect on variability reduction in measuring species richness in both 2023 (4 teams with 2 observers each) and 2025 (16 individual observers) exercises.

The reduction in Jaccard-based dissimilarity confirms the briefing session enhanced agreement in species detection, highlighting its effectiveness in reducing overlooking and misidentification errors. Importantly, a similar reduction in variability was observed in 2023, even without a reference species list, indicating that the interactive briefing session per se was sufficient to promote species familiarization and observer alignment (Vittoz and Guisan 2007).

In contrast, the absence of a consistent reduction in Bray–Curtis dissimilarity indicates that individual differences in abundance estimation persisted despite training, particularly in the 2025 campaign, which involved a larger number of single observers. This pattern was observed even though the 2025 protocol prescribed the use of direct percentage cover estimates, generally considered more robust than Braun-Blanquet classes (Dengler and Dembicz 2023). This indicates that species cover estimation remains a critical source of observer variability, especially for the most abundant species (e.g. Ruscus aculeatus), and may require more specific calibration tools. While plants with low cover are associated with greater observational variability when focusing on species detection (Klimeš et al. 2001; Archaux 2009), in this case it is the cover of abundant species that has the largest impact and potential variation (Vittoz and Guisan 2007). Particular attention must therefore be paid to training observers to arrive at a consensus on abundance estimates by organizing group exercises in which the same species is assessed and discussed simultaneously by all. This exercise can be extended and applied to morphological groups of species (e.g., graminoids such as Carex sp.).

The observed decrease in Euclidean dissimilarity further supports a general convergence among observers, especially in terms of shared detections and non-detections. It has to be stressed again that this distance considers “double zeros” in the same way as “double presences”. Usually, this feature does not allow the use of this index for community comparisons (Orlóci paradox, see Ricotta 2021), but in the context of measuring the differences among observers in detecting species (or in not detecting the same species) it seemed to be effective in mitigating the discrepancies related to plant cover found with the Bray–Curtis dissimilarity. The different patterns observed among dissimilarity indices illustrate how methodological choices may always influence conclusions, in this case about observer agreement, even when their differences are well known and chosen intentionally to answer different questions.

Overall, these results demonstrate that training and intercalibration are effective in reducing observer-induced pseudoturnover related to species presence, while variability in abundance estimation remains an important challenge for future vegetation monitoring programmes. The approach presented in this study addresses the needs of quality assurance for monitoring in forest ecosystems and complies with the requirements and criteria of the ICP Forests Monitoring Programme (Ferretti et al. 2021, section 4.2), but it can be adapted and tailored to the specific type of vegetation being surveyed.

Focusing on key elements of vegetation surveys, we provide some general advice. In programmes covering a wide geographic extent, project planning would assign observers on a geographical basis (e.g., Italian administrative regions). Playing with a common species list is a tool to simulate familiarity with the flora, which is otherwise specific to the survey exercises, providing a baseline to assess remaining biases in terms of species identification and overlooking. Often, the survey protocol requires subdividing the vegetation into vertical layers; this is a potential cause of large discrepancies in cover estimation, therefore the training should focus on correct application of the protocol and all possible edge cases or misinterpretations that could occur. Sampling units differ in both size and structure (e.g., nested or transect) depending on the ecosystem considered, whether it is characterized by denser vegetation with few dominant species (e.g., grassland) or by sparsely distributed species (e.g., shrubland); the training phase should reflect as much as possible all those sampling conditions.

Our results provide evidence that targeted training and intercalibration are effective tools for reducing observer-induced pseudoturnover related to species presence in vegetation surveys, even when observer heterogeneity increases.

Monitoring programmes and resurvey projects should include training and intercalibration exercises as a way to establish a common ground not only in terms of a standard procedure but also of its interpretation, guaranteeing the achievement of programmes’ goals and the expected relative team performances. This is in line with Morrison et al. (2021), who indicated that measures to avoid nonsampling error “can be taken, with or without the use of multiple observers at each site”. These are effective to intercalibrate observers able to adhere to the same standard (e.g. on-field manual and reference species list), regardless of whether they remain the same throughout the survey period or whether they change.

Briefing sessions are a group exercise where observers share their survey experiences allowing them to avoid systematic errors (e.g., protocol misinterpretations) and harmonize judgement (despite personal professional background and knowledge). They consistently improve agreement in species detection, by reducing overlooking and misidentification errors and enhancing the reliability of species richness and presence-absence metrics. In contrast, observer variability in species cover estimation remains largely unresolved, despite the use of direct percentage cover values, indicating that abundance assessment is inherently more subjective and sensitive to observer-specific attributes. These findings highlight that standard training protocols are sufficient to improve detection-related data quality, but not to fully control variability in abundance estimates. Future large-scale vegetation monitoring programmes should therefore complement training with dedicated calibration tools for cover estimation to ensure robust assessments of vegetation change. In the absence of training on plant species cover, we suggest using binary presence/absence data for scientific purposes.