Here we show an analysis of average values of each index, then examine the distributions and finally consider ensemble-averaged maps for each index. For each of these we compare the simulated 1961–90 values with E-OBS, and consider future change in the model projections.
Validation of indices
On average, historical model biases (vs. observed) are small (too short) for Growing Season Length (GSL) and (too late) Start of Field Operations (SFO) (Table 2). By contrast, there are substantial systematic historical model biases in the number of Dry Days (much too few), Air Frosts (too high a degree), and Plant Heat Stress (too large) (Table 2). Biases in the latter two are likely due to both exaggerated model temperature extremes (Ho 2010), and a reduced range of observed temperature values (Haylock et al. 2008). Given that interpolation issues (i.e., underdispersion; Wilks 2011) in the construction of the E-OBS data set (Haylock et al. 2008) are unlikely to affect the order of events throughout the year, it seems likely that late Start of Field Operation figures and mildly short Growing Season Lengths describe modelled winters that are too long and too cold, and modelled summers that are too short.
Rivington et al. (2008b) found that the single member version of HadRM3 generates too many small rainfall events (<0.3 mm), still above our threshold for a “wet day”, compared to the British Atmospheric Data Centre’s daily precipitation record (http://badc.nerc.ac.uk/home/index.html). This is a well known issue with climate models (the “drizzle effect”; Prudhomme 2012) leading to an underestimation of Dry Day occurrence, as seen in Table 2. Haylock et al. (2008) suggests that E-OBS has a significant dry bias for extreme precipitation events above the annual 75th percentile, and Hofstra et al. (2009) describe a significant overall dry bias in E-Obs due to over-smoothing. Sources of bias in our Dry Days index are thus difficult to identify. A higher threshold for the index could help to minimise these issues, but even days with low rainfall (rather than none) are important for agriculture. We have used a lower threshold, but recommend that the reader not overestimate the reliability of low-threshold modelled output for dry-day indices.
Aggregated historical bias and climate sensitivity seem largely uncoupled for our indices. We might expect indices that measure temperature magnitudes (PHS) to behave consistently with a warmer model world, but only when considering average changes in Plant Heat Stress from past to future periods, and the first and last five ranked ensemble members as separate groups (mean values of +25 and +41, respectively), can we say that changes might be linked to sensitivity. Regionally averaged temperature is likely to respond strongly to model climate sensitivity, but measures of relatively uncommon events may be more strongly related to internal climate variability. Given a lack of evident relationship with climate sensitivity, the level of aggregation in Table 2, and the small sample size of this analysis, it appears that the variance of a given ensemble member is more likely to be driven by either model noise or internal variability.
Dry Days is the only index that shows no significant (spatially aggregated) change across ensemble members (Table 2). Modelled Aggregate Air Frosts decline dramatically from their historical values towards 2061–2091, in most cases by between a half and two-thirds. Growing Season Length increases by around a half again, to take up almost three quarters of the year. Plant Heat Stress occurrences more than double in most members, and triple in those members with high climate sensitivities (over 4.5 °C). Start of Field Operations dates shift back into mid-to-late February for the cooler nine members, and early February for the warmest two. These changes are vague by necessity, given that spatial aggregation over the UK removes regional relevance, no single ensemble member is any more likely than any other, and that these are also temporally aggregated averages. They are useful as a rough estimation of changes for the whole of the UK, for the second half of this century, and for a subset of possible futures, but little else. To gain further information we must disaggregate Table 2 into the relevant distributions, or with a greater degree of spatial granularity.
Probability Density Functions (PDFs) are frequently used within climate modelling (e.g. Wilks 2011). Here they allow us to disaggregate bias for a greater level of analysis (Fig. 2). They are used to show not just the ‘location’ (i.e., the mean, Table 2), but also the scale (range) and shape (distribution) of each member against observational data for the past, along-side behavioural shifts for the future. Note that these plots include all years and grid points for the target region over the historical or future periods. This allows us to assess bias in the tails, and whether or not uncertainty may expand or shrink over time through ensemble convergence or divergence. If the PDFs show less (more) agreement into the future, then our sampled futures are diverging (converging).
Both Accumulated Frosts (AF) and Start of Field Operations (SFO) convergence into the future, with a clear tendency toward a lack of future frost (Table 2). Warmer members have a tendency for a narrower distribution and fewer years with many frost days (Fig. 2a). This shift in behaviour is inverted for SFO, describing regimes that are more likely to possess start dates within the first three months of the year. Later (upper tail) dates shift earlier more for warmer members than cooler members. This lengthening of field operations may influence field preparation dates (e.g. tillage) but might also increase the risk of soil compaction if early operations are conducted in wet conditions. Compaction may lead to induced nutrient deficiency, and thus an ultimate yield reduction (Lipiec and Stepniewski 1995). Both AF and SFO show distributions that smooth into unimodal distributions and converge into the future.
Ensemble members all show a mean increase in Growing Season Length of around two months. Distributions of GSL show little convergence for 2061–90 and the future distributions are not well represented by simple shifts of one another (Fig. 2c) By 2061–2090 the coolest three ensemble members and the warmest one are distinct from the rest of the ensemble (smaller and larger increases, respectively, limiting convergence), but most members possess many more years or locations for which the growing season is the entire year. Nearby regions which display a year-long growing season include Portugal and Galicia. Longer growing seasons may allow for a greater diversity of crops (including those with long maturation periods), and the potential for multiple harvests on the same land. Conversely, both irrigation needs and the risk from invasive species, pests and pathogens may increase (EPA 2013).
In all ensemble members (but not observations) Simulated Dry Days (DD) possess a bimodal distribution with a peaks at both 120 and 60 days. This second peak is likely due to overly wet model results for north-western Ireland and western Scotland (Fig. 3). The most likely reason for this under-representation of dry days is a tendency for the model to over-estimate drizzle, as described above. Although the 60-day peak remains relatively static into the future, the 120-day peak increases towards 130 days. Biases appear similar throughout the distribution (as seen in Table 2) and have no apparent relationship with climate sensitivity. In terms of impact, an increase in long-lived dry conditions is also likely to increase the need for irrigation, particularly if compounded with a coterminous increase in Growing Season Length. We have repeated a selection of our Dry Days results for a 1 mm threshold (Online Resource 1), which shows that modelled results are much closer to observations when the daily threshold is increased. However, bimodal behaviour is still apparent and there remains no evident relationship with climate sensitivity. Given that the smaller (100 day) peak is poorly represented in the 1.0 mm plots while the larger (200 day) peak is only slightly underestimated, an excess of drizzle in the model seems likely. Although dry day periods are longer with a larger threshold, there is still little change from the past to future periods.
Many of the Plant Heat Stress (PHS) historical PDFs peak close to zero days, including the observational PDF, but some show a wider range. Future changes are generally towards a broader distribution with no zero-days peak. The future-period PDFs show a tendency for warmer-world members to have a broader spread and greater maxima than the cooler members. There are many locations or years for which PHS does not occur in the past, and this is less evident in the simulations, much less true into the modelled future, and even less so for warm members than cool ones. As suggested above, PHS is more sensitive to changes in model physics than the other indices. Even so, a future in which every year shows some degree of PHS may require complex adaptation strategies, possibly including crop breeds with a greater thermotolerance (Wahid et al. 2007).
Growing Season Length and Accumulated Frosts both shift from a complex, multi-modal distribution towards a unimodal peak, unlike either Dry Days or Plant Heat Stress. Although both Accumulated Frost and Dry Day index distributions narrow into the future, in the former case it is with a shift of mean values towards less frosts, and in the latter, with only a slight change in the mean towards drier conditions. The regional model ensemble (Table 2, Fig. 2) suggests a future UK with fewer frosts (Fig. 2a), fewer years with a large number of frosts (Fig. 2a), and an earlier start to field operations (Fig. 2e). It is a future with more dry days (Fig. 2b), more hot days (Fig. 2d), and a much longer growing season, with 10 and 11 month growing seasons becoming more common toward the end of the century (Fig. 2c).
End-users of climate model data are often most interested in the magnitude of the change that they are likely to experience (see above), types of climate change that are specifically relevant to their needs (e.g. the indices used here) and changes that are specific to their geographical region of interest. Maps of difference over time directly address these three points, and can also provide information about regional change over time.
Maps of Accumulated Airfrost days (Fig. 3a) illustrate how aggregation over a wide area can skew results. The spatial pattern of accumulation is such that although a slight latitudinal effect is apparent, the Scottish Highlands are the main feature in both periods. The Highlands have historically been the region most prone to frost (Fig. 3a), and is projected to see the greatest frost reduction (by in excess of 300 degree days over the year), compared to relatively little change in the south west of England and Ireland (less than +170).
Systematic bias in the number of Dry Days (DD) is also evident in the spatial maps (Fig. 3), as is the double peak behaviour seen in its PDF (Fig. 2). DD increase across Great Britain and Ireland, with the largest increases in the SE, but reductions in parts of coastal NW Scotland (Fig. 3b). The patterns between observational and modelled pasts are consistent, but projected future values (change plus modelled past) are smaller than historical observed values. Stand-alone maps of gridded observational data or model output dependent on rainfall may be suspect, but changes from one period to another might be more reliable. Constant biases may cancel out unless they arise from issues of non-stationarity. Here, DD are shown to increase by over 25 days in the south-east, of England and to decline by 10 in north-western Scotland. As with the PDF plots (Fig. 2), an increase in threshold to 1.0 mm/day (Online Resource 1) provides a much closer fit between modelled and observed results. The change plot for the future period is similar to that for the 0.2 mm/day threshold, although with less intense increases in the south of England, Wales, and Ireland (+15) and a southerly shift in the wetter (−10) band from the coastal N and NW of Scotland to the western and northern Highlands.
For Growing Season Length (Fig. 3) the historical period shows a combination of latitudinal and altitudinal effects, with GSL higher toward the south (around 250 days) and for lower altitudes, than the north and higher altitudes (~190 days). The change map shows a significant increase in GSL across most of the UK (+20–50 days), with less in the highlands (+10–30), and an unexpected decrease in the south-west of England, Ireland, and Wales (−10–30), these changes compliment those in Accumulated Air frost, but the south western decrease is difficult to explain.
PHS values for the modelled past are significantly overestimated in the south of Ireland and England (by ~25–55 days), which displays values at the upper end of the PHS distribution (Fig. 2d), but not in the north. The lower projected Plant Heat Stress values for Scotland (+10–25) may therefore be more reliable than the higher values for the South of England (+40). Although threshold based indices give no indication of changes in the distribution below a given value: There may be a progressive bias where higher values are inflated more than lower values; there may be dynamical reasons for a southern over-estimation; or historical modelled values for Scotland might be ‘right for the wrong reasons’. Were northern PHS values robust, however, an additional three weeks of temperatures in excess of 25 °C would be a significant change.
Although significant shifts in Plant Heat Stress are likely to be the most adaptation-relevant result from this study, there are serious questions as to whether or not the model results are fit for purpose. It is clear that the ensemble (averaged over members) overestimates summer high temperatures, and more so in the south than the north. The projected increase in Plant Heat Stress is almost entirely due to an increase in the average of those temperatures (Online Resource 2a), rather than an increase in temperature variability. Taking average summer temperatures for 1961–1990 adjusted with the mean change (‘delta’) between past and future period summers (3.2 °C) and calculating Plant Heat Stress values from the new constructed future temperatures suggests that the model could be corrected with regards to high temperatures, and that simulated changes may be robust (Online Resource 2b). However, this study does not include information on the timing of heat stress, which may be critical in determining crop productivity, and is an important avenue for further research.
The Start of Field Operations is later in the Scottish Highlands (by 25–50 days) than any other part of the UK during the historical period, with the North of England and Mid-Wales showing SFO values around 80–100 (end of March / beginning of April), 50–70 for the south of England and northern Ireland (end of February), and 15–30 for Cornwall and Southern Ireland (January). Delta values fit this pattern, with later starts becoming proportionally earlier. Although modeled SFO values are slightly biased (toward later values) in Scotland, projections show SFO values around two months earlier in the Highlands. For the rest of the UK, values are a month and a half earlier in the north of England, a month earlier throughout southern England and Northern Ireland and two to three weeks earlier for southern Ireland and Cornwall. Those with earlier starts change less, and the variation in start date is likely to diminish across the UK. Although there is not much agreement between the model ensemble members, there is little systematic bias here, relative to observations, on average.
If we assume that land managers alter their habitats to cope with these new conditions (and specifically days of Plant Heat Stress), earlier and earlier shifts in the start of field operations could imply a generally increased capacity, allowing for earlier fertilization, tillage, and sowing. Recent years have, however, shown that heavy winter or spring rainfall can make fields inaccessible even when warm spring and summer conditions prevail. Model projections show that the wettest regions of the UK, specifically western Wales and western Scotland, are becoming wetter still (Online Resource 3a). When looking at annual totals (Online Resource 3b) we can see that increases in the north-west are more than an order of magnitude greater than in the south east. Ekström et al. (2005) show that the magnitude of long duration, high return period extreme rainfall is declining for most of the UK (−20 %), but increasing substantially for Scotland (+30 %). This may be an issue that keeps Scottish land managers from making the most of extended field and habitat access without sufficient drainage.