Splitting line every x meters using QGIS?

Splitting line every x meters using QGIS?

I have a line that I ultimately want to split up into points. The points should be every 100 meter along the line. So I don't want to extract the nodes.

Are there any Open Source (QGIS, Python) tools around for that?

The use case is that I have a bus line without bus stops. Though I know every 100 meter the bus stops. This way I want to generate bus stops to use as a GTFS feed.

You can use the Convert lines to points tool (you need SAGA GIS installed and the processing toolbox plugin enabled) and set your distance:

This is what I received for my line layer:

I used the Measure Line tool from the toolbar to do a quick check between points:

Hope this helps!

May be an alternative to the suggestion AndreaJ gave- Have look at

Station Lines

N.B. This tool expects projected coordinate systems for the feature to be splitted by planar length unit.

Can be done with OpenJUMP and linear referencing tools. Select the line layer first.

Fill in the parameters

Results go to a new point layer. Points are at equal intervals along the line, original vertices are dropped. Start and end nodes can be preserved if desired.

Frontiers in Marine Science

The editor and reviewers' affiliations are the latest provided on their Loop research profiles and may not reflect their situation at the time of review.


Genetic diversity and structure of the endemic and critically endangered Populus caspica in the Hyrcanian forests

Caspian poplar (Populus caspica Bornm.) is an endemic species from Hyrcanian forests and classified as endangered in Iran. However, there is little information about population genetic structure of remnant populations that could support the development of the suitable conservation strategy for this species. This study was designed to understand how anthropogenic and environmental factors have influenced the pattern of genetic diversity of P. caspica. We sampled 359 individuals from 20 populations. However, after removing the clonal ramets, the genetic parameters were calculated for 314 trees using 14 microsatellite markers. Nine populations showed a significant reduction in effective population size and a genetic bottleneck. The inbreeding coefficient (FIS) ranged from 0.02 to 0.11, and the pairwise FST from 0.018 to 0.224, indicating a low to high level of differentiation among populations. The average proportion of migrants was low (0.008), and it was revealed that the Alborz mountains are a major barrier to gene flow between P. caspica populations in the Hyrcanian forests. Resistance analysis showed that elevation is a significant barrier for gene flow across the eastern and western range of the species. STRUCTURE analysis showed the existence of at least two genetic clusters, whereas GENELAND estimated the number of clusters as three. We propose the maintenance of three natural management units to reduce the risk of extinction of this species in the near future.

This is a preview of subscription content, access via your institution.


Field sites and data collection

Ochrogaster lunifer

Ochrogaster lunifer is a widespread univoltine species found across coastal and inland Australia where its Acacia, Eucalyptus and Corymbia spp. host trees occur [12, 25]. Within the species, there are five nesting types [12] with different ecology, morphology and genetics [26]. In our study, we focused only on the ground nesting form in which larvae create a silken nest at the trunk base of various Acacia and Eucalyptus spp. host trees [18]. The larvae feed throughout the summer until they reach their final instar in autumn, when they leave their nest permanently in search of a suitable pupation site underground. In our study, O. lunifer pre-pupation processions were followed in two seasons, one over seven and the other over ten non-consecutive days from end of March until first week of April 2017 and 2019, respectively. Field visits occurred on other days during the season but there were no active processions. Processions came from various Acacia spp. tree ground nests at The University of Queensland (UQ), Gatton campus, Queensland, Australia (− 27°56′S, 152°34′E, Additional file 2). The campus is a semi-urban environment with buildings and patches of vegetation throughout. Larvae generally left the nest after sunrise, between 06:00 to 07:00 local time. Time at which the leader of the procession left the nest and the time at which the last larva of the procession (or singleton) went underground (bivouac) were recorded (Additional file 3). A bivouac is not necessarily the final site where the larvae pupate, especially for the first day of travel (M. Uemura, personal observation 2017). Most often, the larvae leave the bivouac at the following sunrise to a new bivouac/pupation site. Any disruptions during the procession were recorded, e.g. larvae run over by cars or pedestrians and were classified as human contact. Every time a procession broke into sub-processions or changed direction, coloured flags with codes were used as markers. Once the procession and/or singleton went underground, the GPS coordinates of flags and where caterpillars went underground were recorded with Garmin GPS Etrex10 (in 2017) or iPhone application ViewRanger (in 2019). In-situ environmental temperatures were recorded for O. lunifer populations using Tinytag Plus 2 data-logger TGP-4505 (Hastings Data Loggers, Port Macquarie, Australia) in 2017 but not in 2019. Environmental temperatures for 2019 during the pre-pupation processions were collected from the UQ Gatton station 040082 [27].

Additional file 3 Video MOV Time lapse of a Thaumatopoea pityocampa pre-pupation procession going underground into a bivouac. Video filmed by Mizuki Uemura using an Apple iPhone 8.

Thaumetopoea pityocampa

Thaumetopoea pityocampa larvae are destructive defoliators of pine and cedar trees in the Mediterranean Basin and Southern Europe [28]. Thaumetopoea pityocampa larvae feed throughout winter until spring in northern Italy, when the final instar larvae leave their tent (nest) in the canopy and descend to the ground to find a pupation site [23]. Thaumetopoea pityocampa pre-pupation processions were followed over seven and two non-consecutive days from the first until the third week of April 2018 and last week of March until first week of April 2019, respectively. Field visits occurred on other days during the season but there were no active processions. The field sites were at three pine forests in Veneto, Italy: Monte Garzon (45°30′ N, 11°11′ E, Additional file 4), Precastio (45°31′ N, 11°10′ E) and Carbonari (45°32′ N, 11°10′ E). The selected field sites are planted forest for soil conservation and are also used for recreational purposes such as hiking, cycling and picnics. Processions could not be followed from the tent because they are situated in the canopy of Pinus nigra host trees that can be more than 6 m high. Therefore, procession data could only be collected from the first sighting along the footpath of the field sites, i.e. all T. pityocampa pre-pupation procession data are approximate measures. Once a procession was found and followed until the bivouac, GPS of the bivouac location was recorded using iPhone application ViewRanger and any disruption/movement of the procession were recorded in a similar manner to the O. lunifer processions. Continuous environmental temperatures for T. pityocampa field sites were measured from HOBO Temperature/RH data loggers (Onset Computer Corporation, Macquarie, USA) placed in a village approximately 1 km away from each field site (Tregnago, Veneto, Italy).

Distance, duration and speed of pre-pupation processions

Topographic distance and travel duration by O. lunifer were calculated from processions leaving the nest until completing a bivouac. Once all the larvae from a nest were underground, the distance travelled was measured with a measuring tape or trundle wheel, retracing every movement of the procession. Speed was calculated from the duration and distance travelled from the nest or first sighting to the bivouac. Distance and duration travelled by T. pityocampa pre-pupation processions are an underestimate because it was not feasible to follow processions from T. pityocampa tents that are in the tree canopy. The distance travelled by T. pityocampa was calculated by addition of the height of the nearest host tree with a visible viable nest (estimated from the orientation at first sighting of the procession before it formed a bivouac) and distance from that host tree until the bivouac site. Duration of travel was calculated by dividing the estimated distance travelled by the average speed of T. pityocampa pre-pupation processions. Speed was calculated from measurements of 15 processions on the ground (e.g. travelled x m in x mins).

Orientation of pre-pupation processions

Orientation of every procession is determined by the leader. Each directional change of the procession leaving from the nest to the bivouac was recorded using a handheld compass (in Australia 2017) or iPhone application Compass (in Italy 2018/19 and Australia 2019). For O. lunifer, orientation of all pre-pupation processions leaving the nest and final orientation to the bivouac following the last turn of the leader were used for the data analyses. For T. pityocampa, orientation of processions at first sighting were used for the data analyses.

Light preference of leading larvae

Solar radiation/light (W/m 2 ) was measured for a sub-sample of 2018 T. pityocampa and 2019 O. lunifer processions using the Solar Power Meter (SPM) ISM410 (RS Pro, 2016, London, England). Solar radiation was measured for O. lunifer processions when they left the nest and at final orientation to the bivouac. For T. pityocampa processions, the solar radiation was measured at first sighting. Three solar radiation measurements were taken: directly beside at the same orientation as the leader, 90 o left and right of the leader. Standardised light was calculated by dividing light (W/m 2 ) from the leader’s position by the average of left and right of the leader. Value of 1 means there is no difference between the light intensity at the orientation of the leader/procession and the surrounding. Values more or less than 1 means the leader/procession travelled to the lighter or darker relative to the surrounding environment, respectively.

Risks of pre-pupation processions contacting humans

For O. lunifer, 2019 data for caterpillar contacts with humans were used, because in 2017, some processions were protected from being run over by cars and walked over by pedestrians. Processions for T. pityocampa were not protected from encounters by humans. A human attended area is defined as areas where people travel to and from places by walking, cycling and driving. It was calculated so the two field sites, Australia and Italy can be compared. For O. lunifer, the average human attended area in percent at UQ Gatton campus was calculated by the amount of urban area (e.g. concrete footpath, buildings, roads, etc.) surrounding each host tree at a 10 m radius using QGIS version 3.6.2 Noosa [29] (10 host trees in total). Human attended areas at UQ are not suitable for bivouac/pupation sites because it is made of concrete with the exception of some areas that had leaf litter. For T. pityocampa, the average human attended area in percent was calculated at Monte Garzon by the amount of gravel foot path surrounding each procession at first sighting within a 10 m radius using QGIS (24 processions in total). Human attended areas in Italian field sites are the preferred and suitable bivouac/pupation sites for T. pityocampa because there is exposed dirt/soil and loose gravel (A Battisti, personal communication 2018).

Statistical analyses

All statistical analyses were performed using R Studio version 1.1.419 [30] and an alpha value of P < 0.05 was taken as statistically significant. Mapping was performed using QGIS with satellite images from Google Earth Pro version 7.3.2 [31]. Data for O. lunifer populations collected in 2017 and 2019 were combined in the analyses. Linear models were used to determine if the number of larvae in a procession and environmental temperatures influenced the distance travelled and/or speed of O. lunifer and T. pityocampa pre-pupation processions. To determine if human attended areas affected the distance travelled by O. lunifer processions, a linear model was used, with the variables: distance travelled and human attended area of the host tree where the procession originated (Results in Distance, duration and speed of pre-pupation processions). Distance travelled by T. pityocampa processions were not modelled against human attended areas because the habitat is a pine forest. Procession orientations were represented as rose diagrams made in R Studio using the software package “Circular” [32]. To determine if processions had a preference(s) in orientation, Kuiper’s test of uniformity was used for each rose diagram with the R software package “CircStats” [33]. Each O. lunifer procession was nested within its host tree (10 host trees in total from 2017 and 2019 combined) therefore, Kuiper’s test of uniformity was also used to determine if host trees affected the orientation of processions. Host trees that had more than 10 processions were selected for the Kuiper’s test of uniformity. Light preference of O. lunifer and T. pityocampa pre-pupation processions were analysed using Chi-square Goodness-of-Fit test for a 50:50 distribution. A linear model was used to determine if host tree and number of larvae in the procession affected light preference by O. lunifer and T. pityocampa processions. Host tree could not be tested for light preference in T. pityocampa pre-pupation processions because the exact host trees and timing at which they were at the tree base were unknown.

Material and Methods

Study region

The study region in south-eastern Kenya (1°23′S 38°00′E) is characterized by medium to low precipitation (1079 mm annual average, divided into two distinct short rainy seasons) and moderately hot climatic conditions (mean 21.4°C) (Jaetzold et al. 2007 ). Thus, the region is classified as semiarid. Data were collected along two rivers, Nzeeu River and Kalundu River (Fig. 1). Most of the local people (97%) depend on subsistence agriculture with the cultivation of maize, beans, peas, mangos and pawpaw (Habel et al. 2015b ). The local human population in our study region strongly increased from 1,967,301 (in 1969) to 5,668,123 inhabitants (in 2009) (KNBS, 2012 ). This high demographic pressure caused land-splitting and a lack of agricultural rotation (lack of fallow land stages), accelerating the reduction in soil fertility (Jaetzold et al. 2007 ), so that more (intact) habitats are needed to be transformed into agricultural land.

Land cover data

Historical land cover data were derived from black-white aerial photographs taken in 1961 and 1980 by the British Royal Air Force. Photographs were georeferenced in QGIS ( 2016a ) using polynomial transformation type with nearest neighbour resampling methods in the coordinate reference system Arc 1960 / UTM zone 36S (EPSG: 21036), resulting in pixel sizes of 1.69 m in both directions. Appropriate root mean square errors and digitizing was performed afterwards.

Current land cover data were collected during September 2015 and March 2016, using an Unmanned Aerial Vehicle (DJI Phantom 2 drone), equipped with an orthogonal attached RGB digital camera GoPro HERO 4 Black (GoPro, Inc., San Mateo) mounted on a Zenmuse H3-3D gimbal. Due to limited flight time (about 18 min per mission), we divided aerial surveys into several overflights to cover 13.5 km of Nzeeu River and 17 km of Kalundu River. Flight paths of single overflights were pre-processed with QGIS to guarantee spatial overlapping of about 35–60%. Flights were performed as autopiloted stop-and-turn flights at 40 m flight altitude with 50 m spacing between loops, covering an area of 200 m stream length and 300 m riparian area on each side of the river (equalling 12 ha total area per flight, mean flight length of about 4 km). The attached digital camera was configured using medium resolution settings of seven megapixels and focal length of 21.9 mm equivalent, resulting in picture dimensions of 2250–3000 px, with aspect ratio of 3–4 reduce fish-eye distortion. Pictures were taken with a 2 sec time-lapse interval, producing an average of about 400 pictures per flight. Additionally, coloured markers were used as ground control points (GCP), which were measured using a Garmin eTrex 10 GPS device. The aerial photographs were combined afterwards with the AgiSoft Photoscan Professional software (Agisoft 2015) using medium-quality dense cloud processing and mesh construction settings. Based on sufficient ground control points, processed imagery was exported as orthomosaic into geotif raster files with geometric accuracy below 1.97 m (1.00 m in longitudinal error, 1.38 m latitudinal error and 0.99 m altitudinal error). The tiled orthophotos were subsequently mosaicked using gdal-function merge in QGIS (GDAL 2015) and prepared for further analysis. For the quantification of land cover changes, we distinguish between the following three land cover categories: (1) riparian thickets (pristine as well as L. camara-dominated thickets), (2) agricultural land (1 & 2 both as polygons) and (3) human settlements (as points, with each point representing an aggregation of several houses). Digitizing was done with the software QGIS Development Team ( 2016a ).

Land cover change

We assessed land cover changes for three spatial scales and with three methods: First, we analysed land cover changes for a 1.3 km section of Nzeeu River (Fig 1). Land cover changes were recorded within a 100 m strip on each side of this river, for three time windows: 1961, 1980 and 2015/2016. For this analysis, we also assessed all human settlements to quantify the effect of demographic pressure. Here, the strip mapped was extended to 300 m as most people do not settle directly along the river due to periodic flooding (Fig. 2). Due to the fact that L. camara invaded major parts of Kenya during the 1950s and later (further details: see Introduction), we assume that all riparian vegetation detectable in the year 1961 represented intact pristine forests.

Second, to exclude potential local bias of land cover change at the above described local scale, we randomly set 50 points (25 along Nzeeu River, 25 along Kalundu River) for the years 1961 and 2015/2016 (with identical points for the two time cohorts, allowing temporal comparative analyses) (Fig 1). Each point was the centre of a 100 m radius plot (i.e. surface area 3.14 ha). These points were randomly selected using the QGIS function ‘random points along line’ (QGIS, 2016b ). Here, we used the same land cover categories as applied in our first method.

Third, due to the fact that the proportion of L. camara thickets was not distinguishable from the pristine riparian forest coverage based on aerial photographs, we assessed the land cover categories for a 50 m strip along the two rivers, Nzeeu and Kalundu, with a hand-held Garmin eTrex 10 GPS device. In this third analysis, we further distinguished the land cover category ‘riparian vegetation’ as two sub-categories: ‘pristine riparian forests’ and ‘disturbed riparian thickets dominated by L. camara’ (with > 50% of the vegetation consisting of L. camara).

Statistical analysis

Analyses of land cover changes were calculated with QGIS and the MOLUSCE tool 3.0.11 ( The MOLUSCE tool provides statistics for land cover class and transition matrix for the time steps to be compared. As MOLUSCE requires raster datasets as input data, we transformed the digitized land cover vector datasets into raster datasets with cell sizes of 1 m, using the QGIS conversion function. Relative errors from rasterization were calculated according to Liao and Bai ( 2010 ).

For both the local and landscape scale (steps 1 and 2, see above), we calculated land cover changes incorporating all categories. Here, we particularly focused on the conversion from pristine riparian thickets into agricultural land, the partial transformation into fallow land and subsequent succession by L. camara. Due to the fact that L. camara was first recorded in Kenya in the wild during the 1950s (Cilliers and Neser 1991 ), we assumed that all riparian thickets detectable on historical aerial pictures from 1961 consist of pristine and diverse plant species composition.

We applied Land Use and Land Cover Change (LUCC) analysis to quantify the area of land cover changes. Changes were statistically validated by transition matrix, simulation and with the help of kappa statistics, which measure the agreement of two categorical datasets on a cell by cell comparison (Van Vliet et al. 2011 ). We considered the limitations of kappa statistics (cf. Van Vliet et al. 2011 ) and accordingly extended our calculations using Kappa histogram and Kappa location to clearly distinguish between quantification and location errors or agreement. Kappa histogram measures the expected agreement between two datasets based on the distribution of class sizes Kappa location measures the maximum agreement given by the distribution of class sizes (Van Vliet et al. 2011 ).

Although the output of the MOLUSCE tool especially concentrates on spatial changes between time cohorts, we extended the analysis with the help of landscape metrics calculated with FRAGSTAT version 4.2 to gain insights into temporal changes of the land coverage between time cohorts. Here, we used an exhaustive sampling strategy for landscape scale using user-defined tiles, which were congruent with our sampling points for the LUCC analysis. At the landscape scale, we used the 25 randomly set and non-overlapping circular plots for each river (with 100 m radius) (see above). We calculated the following two indices on class and landscape level: (1) number of patches at class level and landscape level. This index provides the total number of patches for a specific landscape (i.e. the respective points of time) and classes (land cover categories). (2) Effective Mesh Size (MESH) which provides area information on fragmentation of the study area. MESH was chosen because it takes patch size and the distribution of a land cover category as well as the total landscape area into consideration (the lower limit of MESH is constrained by the cell size and is achieved when the landscape is maximally subdivided, that is, when every cell is a separate patch MESH is maximum when the landscape consists of a single patch). Results from FRAGSTAT were post-processed using SPSS version 22 including Kruskal–Wallis H and Mann–Whitney U tests to compare the distribution of the land cover category and structural changes between the two rivers and time cohorts.

Arthropod abundance

Aggregated abundances of major arthropod orders and families were assessed using 200 pitfall traps along Nzeeu River (100 traps) and Kalundu River (100 traps). Five traps were buried in a star-wise pattern, with one at each cardinal point and a central trap, with each trap at least 5 m distant from its neighbour. Collected material from such a pitfall site (consisting of five traps) was afterwards merged and treated as coming from one site. Hence, we obtained material from 20 sites (10 sets in L. camara-dominated thickets, with more than 50% thicket consisting of L. camara, and 10 sets in pristine riparian forests, still not invaded by L. camara), from each of the two rivers. Each pitfall trap, with a diameter of 9 cm, was buried in the soil with its top at ground level, and filled with saturated salt solution. Traps were left in the field for 5 days (at dry weather conditions). Collection was repeated in a second round (another 5 days) under identical conditions. Sampling was conducted in March 2014 and 2015.

Arthropods were split into three groups: mainly herbivores (i.e. Lepidoptera, Homoptera, Hemiptera, Orthoptera, Cicadina), mainly predators (i.e. Arachnida, Chilopoda, Carabidae, Mantodea, Neuroptera, Staphylinidae), and ants. For each group, the average abundance between both habitat types (L. camara-dominated thickets vs. pristine riparian forest) was compared using non-parametric one- and two-way PERMANOVA (Bray-Curtis dissimilarities, 5000 randomizations performed in Primer 7.0) using study year and habitat as grouping variables.

3. Methods

In this section, we outline the methods used to assess seasonal forecasts, consistent with the concept of forecast “goodness” as discussed by Murphy (1993). Various metrics exist to measure the quality or performance of a forecast, and we focus on skill, reliability and bias. Skill considers how the forecast compares to a reference forecast, for example a forecast of no skill (“random chance”) or a persistence forecast. Reliability quantifies the probabilistic agreement between the forecasts and the observations. Bias quantifies the extent of any systematic differences, on average, between the forecasts and the observations.

In our analysis we concentrate on the Sahel area, defined as 10°–20°N, 20°W–30°E (marked as a green box in Fig. 3a). This area is similar to that used by Vellinga et al. (2016) but extended farther west to include all of Senegal. We evaluate the forecasts for different time periods, owing to differences in data availability, ensuring we use concurrent data for comparisons between the dynamical model forecasts and PRESASS forecasts. In section 4a, we consider the period of overlap between GloSea5 hindcasts and the PRESASS forecasts (1998–2015), and in section 4b we use the period of overlap between the different models hindcasts (1993–2010).

A. Digitization of PRESASS forecasts

The digitization was completed using QGIS (QGIS Development Team 2018), with more detailed instructions available from the ASPIRE website (Met Office 2019). When the 20-yr forecasts are examined, a natural development of the forecasting process is apparent, with two key implications. First, the forecast area is not consistent, so we only use forecasts from the 18 countries shown in Fig. 1 excluding the Cape Verde Islands. This represents the main core of the countries that appears on most of the forecasts throughout the 20 years.

Second, there is not a consistent way to interpret the area shown in gray in Fig. 1. Analyzing the 20 forecasts Figs. S3 and S4 shows two different approaches, which are sometimes colored differently (e.g., 2012) but sometimes the same (e.g., 2017). In the north, the climate is arid so in most years, no forecast is made. Conversely, farther south and more recently, gray becomes a forecast of “near-average precipitation,” but there is no boundary between the regions for no forecast due to aridity and forecasts of near-normal rainfall (except in 2012 and 2013). However, the boundary between the regions of no-forecast and implicit near-normal forecast varies across the 20 forecasts, so it is not possible to identify an area where null forecasts can be consistently assumed. This raises the question of what percentage probability should be assigned to the gray or white areas. An even split across all three terciles (33–33–33) would represent a null forecast and is a reasonable assumption made by Walker et al. (2019), but this would not be a forecast of near-normal precipitation and is forecast explicitly in 2006. A near-normal precipitation forecast could be represented by percentages that are skewed toward the middle tercile such as 30–40–30, but this is forecast explicitly in 2004. These conflicting motivations for having regions with no explicit forecast and different interpretations means that in this study, we use only the regions where the PRESASS process explicitly gives a percentage forecast.

B. Skill assessment

As an initial assessment, correlation maps illustrate the spatial characteristics of deterministic skill. At each point, the Pearson correlation coefficient is calculated between every year’s ensemble mean forecast and the observations (Fig. 3). Where the correlation is statistically significant (at the 95% confidence level), the map is hatched. The Pearson correlation is chosen for its simplicity and common usage, but relies on normally distributed data. While this is not the case for daily precipitation, calculating season total precipitation means the values tend to become normally distributed but does not guarantee that the correlation is unaffected by outliers. For this reason, we also calculated the Kendall tau correlation, based on ranking and not affected by outliers, and the results are qualitatively similar (not shown) to those for Pearson correlation.

Taking a probabilistic approach, relative operating characteristic (ROC) scores assess the skill of forecasts in terms of whether a forecast “event” occurs. Events are expressed as precipitation falling into each tercile (above, near, and below normal) in turn. Thresholds are applied to the forecast probability for each event, with the event being forecast if the probability exceeds the threshold. The corresponding point in the observational data is identified, with four possible outcomes:

  • hit: the event is forecast and occurs
  • false alarm: the event is forecast and does not occur
  • miss: the event is not forecast and does occur and
  • correct negative: the event is not forecast and does not occur.

In this paper, ROC diagrams are used to illustrate the skill of seasonal forecasts produced using different models over the Sahel. The vertical axis of a ROC diagram shows the “hit rate,” which is the number of hits compared to the total number of observed occurrences (number of hits plus the number of misses). The horizontal axis shows the “false alarm rate,” which is the number of false alarms compared to the total number of nonoccurrences (the number of false alarms plus the number of correct negatives). For each tercile of precipitation, these two rates are plotted against each other for each forecast probability threshold. The resulting curve shows whether or not the forecast system is l. A skillful forecast system will maximize the hit rate and minimize the false alarm rate, so the ROC curve would bow toward the top left of the plot. If the forecast system has no skill, then the false alarm rate and hit rate would be equal, with the ROC curve along the straight diagonal line. As such, ROC diagrams indicate the forecast skill compared to a “random chance” forecast. A ROC skill score (ROCSS) is given in the plot legend (Fig. 4), calculated as per Wilks [2011, Eq. (8.46)]. ROCSS is 1 for perfect forecasts and 0 for “random guess” forecasts.

In addition, ROC maps can be used to provide spatial illustrations of forecast skill. The ROC maps used in this study show the area under a ROC curve at each grid point, for each precipitation tercile separately. A forecast with no skill, and ROC curve that falls along the diagonal, equates to an area under the curve of 0.5. Therefore, where a forecast has skill, the ROC map shows a value of greater than 0.5 the higher the value, the more skillful the forecast. Note that the area is calculated using the trapezoidal rule, which will slightly underestimate the area under the curve so should be used primarily to compare different areas of the forecast and the different forecast systems.

C. Reliability assessment

Using the example PRESASS forecast in Fig. 1, and focusing on the orange forecast area, there is a 50% probability of the area receiving below-average precipitation, a 30% probability of receiving near-average conditions and a 20% probability of above-average conditions. In the scenario that this region received a perfectly reliable forecast, we would expect 20% of the points in this region to receive precipitation in the above-normal tercile, 30% near normal, and 50% below normal. This is repeated over the 20 years, in order to fairly assess the reliability.

To assess the reliability, we collected forecast probabilities for each tercile over the 20 years of available forecasts. For example, we collected the cases where the forecast of above-average precipitation was 20%, and count how many times that the precipitation was in the above average tercile if this occurred 20% of the time, then those forecasts would be reliable. We assess the forecasts in intervals of 10% probability.

Reliability diagrams illustrate how reliable forecasts are by comparing forecast probabilities with the frequencies of actual events. The vertical axis of a reliability diagram shows the observed frequency, and the horizontal axis shows the forecast probability. In this study we group the observed frequency and forecast probabilities into bins with 10% intervals. The lower half of Fig. 2 shows two example diagrams. As in the example described above, reliable forecasts give similar values for forecast probability and observed frequency, so perfect reliability would result in a line on the main diagonal (y = x).

(top) ROC diagrams and (bottom) reliability diagrams for (left) PRESASS forecasts and (right) GloSea5 hindcasts of July–September season total precipitation for 1998–2015.

Citation: Weather and Forecasting 35, 3 10.1175/WAF-D-19-0168.1

D. Bias assessment

Frequency diagrams (or sharpness diagrams) illustrate the forecast bias. Examples are shown at the bottom of Fig. 2, giving the number of forecasts within each bin over all years and for the Sahel, our region of interest. Frequency diagrams are used in two ways: first, to ensure there are a sufficient number of forecasts in each percentage category to allow statistically significant interpretation and second, to assess probabilistic bias.

Over a long period of time, the frequency of forecast probabilities in each tercile will converge to the climatological frequency of one-third (0.33). If on average a tercile is forecast with a higher (or lower) percentage chance than the climatological average, this will be represented as a shift in the histogram away from a center point of 0.33 and indicate a probabilistic bias in the forecast. Note that because the model forecasts are assessed against terciles calculated from model climatology, they will not exhibit probabilistic bias, and we do not assess absolute bias (i.e., the long-term average difference between model hindcast mean and observed mean).


We thank the SERC for logistical support, William Brinley and Nathan Phillips for technical assistance in construction of the sapflow sensors and Geoffrey Parker for providing access to the 50-ha plot and for other support. We thank Marilyn Fogel for allowing us to use her former facilities the Geophysical Lab at the Carnegie Institution in Washington, DC for isotopic analyses, and Lauren Urgenson, George Raspberry (posthumously), Roxane Bowden, Kati Dawson Andrea Krystan, and Patrick Neale for technical and other assistance. The water table and soil moisture data were gathered as part of an NSF funded project to Sean McMahon (NSF Grant 1137366) and is curated by Rutuja Chitra-Tarak.

4.2 Ocean mean state and identified biases

The annual SST averaged between 50 ∘ S and 50 ∘ N simulated by IPSL-CM5A2 for the last decades of the historical run is 22.3 ∘ C (1 ∘ C warmer than IPSL-CM5A), a value in the higher range of CMIP5 ESMs and consistent with observations (Fig. A2).

Figure 10Yearly averaged anomalies between simulated surface temperature (a, b) and surface salinity (c, d) and WOA2013 observations (Locarnini et al., 2013 Zweng et al., 2013) for the 1980–1999 period. (a, c) IPSL-CM5A – WOA2013. (b, d) IPSL-CM5A2 – WOA2013.

The warm bias over eastern subtropical Pacific and Atlantic oceans is more marked in IPSL-CM5A2 than in IPSL-CM5A as a response to CRF tuning (Fig. 10a, b). This bias has already been reported in several coupled models using NEMO, namely CNRM-ESM1 (Séférian et al., 2016) , CNRM-CM5.1 (Voldoire et al., 2013) , IPSL-CM4 (Marti et al., 2010) and IPSL-CM5A (Dufresne et al., 2013) . As stated in Voldoire et al. (2013) , ocean-only experiments (Griffies et al., 2009) have suggested that such a warm bias was likely linked to “poorly resolved coastal upwellings and underestimated associated westward mass transport due to the coarse model grid resolution”. The 0.5–2 ∘ C warm bias over the Antarctic circumpolar current (ACC) is comparable to IPSL-CM5A and can be related to the poor simulation of the CRF over this region (Hyder et al., 2018) and the associated mispositioning of midlatitude westerlies depicted above. Conversely, the tuning has reduced the cold bias in the Pacific and Atlantic gyres. Surface salinity biases are rather similar between both models. The surface salinity anomaly follows the evaporation minus precipitation (E–P) pattern over the tropical oceans (not shown), showing the influence of the double ITCZ issue on the tropical ocean freshwater balance. Both models depict too-fresh surface waters in the northern Atlantic and locations of deep water formation (Fig. 10c, d). The ocean also depicts temperature biases at depth both in the Atlantic and Pacific oceans (Fig. 11). Zonally averaged potential temperature anomalies show that the warm bias in the northern Atlantic between 1000 and 3500 m has been amplified, as well as the warm bias in the northern Pacific, that concerns the entire water column in IPSL-CM5A2. The strong ( > + 2.5 ∘ C) overestimation of temperatures in the subsurface waters of the Southern Hemisphere is almost identical in both versions of the model. The cold ( ∼ - 1 ∘ C) bias in the high latitudes of the Southern Hemisphere has been slightly reduced but is still present due to the lack of strong enough overturning in this region.

Figure 11Longitude–depth cross section of yearly averaged anomalies between IPSL-CM5A2 temperature (a) and salinity (b) and WOA2013 observations (Locarnini et al., 2013 Zweng et al., 2013) averaged over the 1980–1999 period.

Sea-ice extent (Fig. 12) has been improved in the North Atlantic sector but remains overestimated. The Nordic Seas are covered by sea ice in winter in IPSL-CM5A, while IPSL-CM5A2 shows overestimated sea ice mostly in the Barents Sea. In the Labrador Sea, both IPSL-CM5A and IPSL-CM5A2 have overestimated sea ice, likely preventing water mass sinking and convection in this region. In contrast, sea ice is underestimated in both versions for the North Pacific sector, i.e., in the Bering and Okhotsk seas.

Figure 12(a, b) March sea-ice cover (%) for the IPSL-CM5A (a, c) and IPSL-CM5A2 (b, d) historical runs. The black contour indicates the 15 % sea-ice cover interval in observations. (c, d) January–February–March average of mixed-layer depth (meters) for IPSL-CM5A (a, c) and IPSL-CM5A2 (b, d).

4.2.1 Ocean heat transport and meridional overturning circulation

IPSL-CM5A2 ocean meridional heat transport is higher (in absolute value) than the one simulated with IPSL-CM5A in both hemispheres (Fig. 13). This increase in the Northern Hemisphere is mainly related to the increase in the Atlantic meridional heat transport in the midlatitudes (not shown). When compared to observations, the two model versions have asymmetric biases, with the northward transport being underestimated in the Northern Hemisphere and the southward transport overestimated in the Southern Hemisphere low latitudes (between the Equator and 35 ∘ S). Southward heat transport is too strong between the Equator and 40 ∘ S and is in better agreement with the data than IPSL-CM5A between 40 and 60 ∘ S.

Figure 13Global meridional heat transport in the ocean, in petawatts (PW).

The AMOC is more vigorous in IPSL-CM5A2, with a maximum in depth (below 500 m) and latitude (from 30 ∘ S to 60 ∘ N) in pre-industrial conditions moving from 10.3±1.2 Sv (Escudier et al., 2013) to 12.02±1.1 Sv (taken for the last 1000 years of simulation Fig. 14a). The last 20 years of the 20th century in the historical run depict a less vigorous circulation in the Atlantic, with a mean value for the AMOC index of 10.5±0.9 Sv.

Figure 14Atlantic meridional streamfunction (a) and global barotropic streamfunction (b) for the pre-industrial experiments at steady state (100-year averages). Color shading shows the IPSL-CM5A2 minus IPSL-CM5A anomaly. Contour lines shows the IPSL-CM5A2 values (in Sv).

The convection sites have been slightly modified between the two versions of the model. While the strongest convection was located in the North Atlantic subpolar gyre (SPG), south of Iceland, in the IPSL-CM5A version (Escudier et al., 2013) , in IPSL-CM5A2, the Greenland Sea shows the deepest mixed-layer depth, which is larger than 1500 m in winter (Fig. 12c, d). Nevertheless, there is still a very active deep convection site south of Iceland, which is not realistic. The strengthening of the Nordic Seas convection can be related with the decrease of its sea-ice cover in winter (Fig. 12). As a consequence, denser water is now produced in this site, which improves the overflow through Greenland–Iceland–Scotland straits. Indeed, the water denser than 27.8 kg m −3 flowing southward reaches 1.5 Sv in IPSL-CM5A2, while it was only 0.2 Sv in IPSL-CM5A. It remains largely underestimated as the observations from Olsen et al. (2008) indicate around 6 Sv of overflow in total from observation-based estimates over the last few decades. Besides, the poleward shift of the westerlies in the North Atlantic stimulate a northward shift of the boundary between the Atlantic subtropical and subpolar gyres, as shown by the barotropic streamfunction anomaly (Fig. 14b). This may contribute to bringing more salt into the subpolar gyre, which may also act to intensify the AMOC in IPSL-CM5A2.

Table 3Mass fluxes through major ocean gateways in Sverdrups (Sv).

4.2.2 Fluxes in the major ocean gateways

We computed oceanic mass transport through selected longitudinal and latitudinal cross sections corresponding to the major gateways (Table 3) and compared them to IPSL-CM5A, estimates from observations and values simulated by CNRM-CM5.1, which use NEMO as well (Voldoire et al., 2013) . These diagnostics highlight that the ACC is underestimated by IPSL-CM5A2 by around 25 %, as for the other configuration considered here, whereas despite the absence of high resolution and explicit resolution of Indonesian gateways, the overall Indonesian Throughflow tends to be close to observation-based estimate (underestimation of 16 %). Compared to observations and CNRM-CM5.1, the flow in the Florida Straits is quite weak (weaker than observation by 37 %), in line with the simulated “sluggish” AMOC. The increase of the AMOC from IPSL-CM5A to IPSL-CM5A2 has not been enough to strengthen this current, also largely resulting from wind forcing. The absence of resolved eddies can also play a significant role in this underestimation.

4.2.3 Net primary production (NPP)

To illustrate some of the capabilities of IPSL-CM5A2, we also evaluate the simulated oceanic NPP (Fig. 15). The global integrated NPP over the historical period (here averaged over 1980–1999 from the historical simulation) is 47.5 PgC yr −1 , which is a strong increase when compared to 30.9 PgC yr −1 simulated for IPSL-CM5A (Bopp et al., 2013) , making IPSL-CM5A2 closer to the 52.1 PgC yr −1 global estimate based on SeaWiFS remote-sensing observations (Behrenfeld and Falkowski, 1997) . The representation of NPP at the spatial scale compares well to observation-based estimates for the tropical and Southern Hemisphere but depicts large biases in the Northern Hemisphere, especially in the North Atlantic where NPP is largely underestimated (Fig. 15c).

Figure 15Mean vertically integrated NPP retrieved from SeaWiFS observations (a) and simulated by IPSL-CM5A2 (b). (c) Latitudinal variations of mean vertically integrated NPP, zonally averaged, for SeaWiFS (black), IPSL-CM5A (orange) and IPSL-CM5A2 (blue).

4.2.4 Ocean variability

The main modes of variability of the AMOC are analyzed with an empirical orthogonal function (EOF) of the yearly Atlantic meridional overturning streamfunction (Fig. 16, top). The first EOF is a monopole in IPSL-CM5A and IPSL-CM5A2, but IPSL-CM5A2 shows less loading in the subpolar gyre between 40 and 60 ∘ N. The associated principal component shows a significant 20-year variability in the case of IPSL-CM5A. Such variability was previously linked to advection of surface salinity anomalies from the western boundary to the eastern Atlantic (Escudier et al., 2013) and westward-propagating planetary waves in subsurface (Ortega et al., 2015) . North Atlantic observations (Vianna and Menezes, 2013 Swingedouw et al., 2013) and proxy records indicate a 20-year preferential variability in this region in the atmosphere (Chylek et al., 2011) , sea ice (Divine and Dick, 2006) and the ocean (Sicre et al., 2008 Cronin et al., 2014) . However, the peak of variability at 20-year timescale is not significant any more in IPSL-CM5A2 (see spectra in Fig. 16c–d). This difference can be explained by the different tuning of the two model versions.

Figure 16(a, b) Spatial patterns of the leading EOF of the yearly AMOC computed from 2490 years of the IPSL-CM5A2 PREIND (model years 4510 to 6999) and from 1900 model years of the IPSL-CM5A PREIND (model years 1800 to 3699), respectively. The patterns are given in Sv for 1 standard deviation of the corresponding principal component, and the amount of variance explained by the pattern is given in the top right of each panel. Panels (c) and (d) show the variance spectra of the associated leading principal components, respectively. The dashed lines show the 95 % and 99 % confidence levels of the best-fit first-order Markov red noise spectrum, respectively. (e, f) Atlantic multidecadal variability (AMV) pattern, in K, in IPSL-CM5A2 PREIND and IPSL-CM5A PREIND, respectively, from the same model years as above. The AMV was calculated from the low-pass-filtered (61-month running mean) North Atlantic SST in 0–60 ∘ N, 80 ∘ W–0 ∘ E. The AMV pattern is given by the regression of the SST onto the AMV index.

Gastineau et al. (2018) used IPSL-CM5A-MR, a variant of IPSL-CM5A with an improved horizontal resolution in the atmosphere ( 2.5 ∘ × 1.25 ∘ vs. 3.75 ∘ × 1.87 ∘ ) and also found that a warmer mean state led to a reduction of the 20-year variability in IPSL-CM5A-MR. Such a warmer mean state modifies the subpolar gyre currents and the subsurface stratification in the eastern Atlantic Ocean where the mixed layer is deepest in IPSL-CM5A models. These changes were suggested to explain a smaller growth of baroclinic instabilities triggering the propagation of westward-propagating anomalies (Gastineau et al., 2018) . It is therefore likely that the different tuning of IPSL-CM5A2 led to similar changes. The SST associated with the first principal component of the AMOC also shows weaker SST anomalies in the subpolar gyre when compared to IPSL-CM5A (not shown). The AMV patterns are similar in IPSL-CM5A and IPSL-CM5A2, with a slightly more intense subpolar gyre SST (6.8 ∘ C vs. 6.2 ∘ C for the maximum value in the subpolar gyre) in IPSL-CM5A2. Larger positive SST anomalies in IPSL-CM5A2 are also simulated in the Nordic Seas and Irminger Sea associated with the AMV. The larger AMV in IPSL-CM5A2 is consistent with the smaller sea-ice cover in the North Atlantic and a more intense signature of the atmospheric forcing. Yet, the observed AMV pattern computed from the historical period (1870–2008) using a similar methodology does not show such an intense subpolar pole: the SST anomalies in the subpolar region are only twice as intense is the tropical ones (Deser et al., 2009) .

The Pacific variability in IPSL-CM5A2 is identical to that found in IPSL-CM5A. The El Niño–Southern Oscillation (ENSO) is characterized by the NINO3.4 time series. IPSL-CM5A and IPSL-CM5A2 show a similar power spectrum (Fig. B3a, c, e) which is comparable to observations (Bellenger et al., 2014) . Both versions have the same bias in the seasonal phase locking (Fig. B3b, d, f), as the monthly standard deviations are larger in May and June in CM5A and CM5A2, while the observed maximum is in January and December. The ENSO impacts are evaluated using spatial composites of the SST over ocean, 2 m air temperature over land and sea-level pressure. The composites are built using the NINO3.4 time series and calculating the difference of all years where the NINO3.4 exceeds 1 standard deviation, minus years less than 1 standard deviation. Figure B4 illustrates that the ENSO impacts are identical in both versions of the model, with a systematic bias in the position of the Pacific equatorial SLP anomalies which extend too far west toward the Pacific warm pool. However, the SLP anomalies display a realistic pattern in the Northern Hemisphere and Southern Hemisphere, with the associated Pacific North American and Pacific South American patterns, as in other CMIP5 models (Weare, 2013) . The Pacific Decadal Oscillation (PDO) was also calculated as the first EOF of the monthly Pacific SST anomalies north of 20 ∘ N (Fig. B5), which illustrate a pattern with positive SST anomalies over the equatorial Pacific, negative anomalies in the northwest Pacific and a horseshoe pattern of positive SST anomalies in the Northern Hemisphere. The IPSL-CM5A model generally simulates the characteristics of the observed PDO, as found by Fleming and Anchukaitis (2016) or Nidheesh et al. (2017) , as well as the associated anomalies in the Atlantic and ocean basins. The IPSL-CM5A2 version simulates a PDO nearly identical to that of IPSL-CM5A, so that the mean state changes have no impact on the PDO state.

This section depicts both the technical developments and the choices to be made regarding boundary conditions for a typical deep-time simulation with IPSL-CM5A2. We describe the generation of a new grid for the ocean model for deep-time applications and the generation of boundary conditions, component-wise. To illustrate our point, we run a 3000-year long numerical simulation of the Cretaceous as a case study. We present a very brief analysis of the model time series to illustrate the long adjustment required to reach equilibrium in deep-time simulations. As the focus of the paper is the description of IPSL-CM5A2, we only provide a very basic description of the simulated climate. We refer readers to Laugié et al. (2020) for a detailed description of a Cretaceous simulation run with IPSL-CM5A2.

Figure 17Comparisons of ORCA2 grid (a) and PALEORCA grid (b). Color shading shows spatial resolution (km 2 ). Grey-shaded areas on ORCA surround the refinements of the grid over the Mediterranean, Red, Black and Caspian seas. The centers of the blue areas show the two poles that have been rotated in PALEORCA.


The world is experiencing a growing abandonment process of previously intensively managed agricultural lands (Baldock et al. 1996 Silver et al. 2000 Dunjó et al. 2003 Benayas et al. 2007 Bergen et al. 2008 Zhang et al. 2010a, 2010b Prishchepov et al. 2012 Alcantara et al. 2013 Hansen et al. 2013 Margono et al. 2014). Many investigations have been devoted to the dynamics of soil and vegetation on abandoned lands, features of the spatial structure of their vegetation, changes in their biodiversity, in carbon and nitrogen budgets, and to peculiarities of carbon sequestration at different stages of forest recovery (Gough and Marrs 1990 Silver et al. 2000 Baniya et al. 2009 Tokavchuk 2010 Zhang et al. 2010a De Frenne et al. 2011 Heikkinen et al. 2014 Hou and Fu 2014 Kurganova et al. 2014 Nyawira et al. 2016 Arévalo et al. 2017 Baeva et al. 2017 Telesnina et al. 2017 Kalinina et al. 2018).

The area and proportion of abandoned lands in the Russian Federation, especially in its European part, are really huge (Lyuri et al. 2010 Kurganova et al. 2014 Potapov et al. 2015). Lands were massively abandoned in the late 1980s—early 1990s due to socio-economic and institutional changes that occurred in the country after the collapse of the Soviet Union. Between 1990 and 2007, 45.5 million ha of agricultural lands were abandoned over entire Russia of which 32.6 million ha in its European part (Lyuri et al. 2010 Kurganova et al. 2014). The situation varies across regions: in the Yaroslavl region lands removed from agricultural usage account for 418 thousand ha (43% of the total area of agricultural lands in the region before the collapse of the Soviet Union Gulbe 2009) and in the Smolensk, Kaluga, Ryazan, Vladimir, and Tula regions such lands account for 46, 30, 28, 27, and 26%, respectively (altogether 1.7 million ha Prishchepov et al. 2013) in the Leningrad and Sverdlovsk regions, these values are 90 and 14% (632 and 660 thousand ha, respectively Novoselova et al. 2016). In European Russia, lands are more likely to be abandoned if they are situated further away from populated localities and market centers, or lie at forest edges or within forest tracts furthermore, a higher likelihood of agricultural land abandonment is found to be significantly correlated with lower average grain yields in the late 1980s (Prishchepov et al. 2013). The specific conditions of these abandoned lands, such as their spatial parameters, soil characteristics, soil seed banks, neighboring plant communities, the type of previous farming, external impacts, etc., are rather diverse (Lyuri et al. 2010). It has been shown that all these parameters, in various combinations, define the trajectories of vegetation succession and features of vegetation recovery on the abandoned lands (Baniya et al. 2009 Baeten et al. 2010 De Frenne et al. 2011 Munroe et al. 2013 Hou et al. 2014 Plieninger et al. 2014 Kou et al. 2016 Stuble et al. 2017).

For Central European Russia, it was shown that pioneer trees, such as European white birch (Betula pendula), gray alder (Alnus incana), or goat willow (Salix caprea), can settle on abandoned arable lands within the first 2 years, and in very high densities (Utkin et al. 2002, 2005). Intensive natural thinning of such stands occurs especially between 4 to 10 years of age while the density of the stands remains relatively high: up to 110,000 ind/ha were counted in 5-year-old Alnus incana stands (Gulbe 2012) and up to 66,000 ind/ha in 10-year-old Betula pendula stands (Gulbe 2009). Further, 30-year-old Betula pendula forest developed on formerly plowed fields located within an old-growth multi-species broad-laved forest in the Kaluzhskie zaseki Reserve held between 500 to 1725 ind/ha in the overstorey (Smirnova et al. 2017). In this 30-year-old birch forest, in a 100-m-wide zone along the old-growth forest margin, shade-tolerant trees predominated in the understory: the total number of tree individuals varied from 20,558 to 83,314 ind/ha European ash (Fraxinus excelsior) and Norway maple (Acer platanoides) dominated. European aspen (Populus tremula), Scots pine (Pinus sylvestris), and Norway spruce (Picea abies), as well as birch and goat willow, also occurred in former fields located in the middle and south taiga (Novoselova et al. 2016). Spontaneous regrowth of forest vegetation occurs even in the fields that have been in agricultural use for a long period of time (up to several centuries and longer) (Utkin et al. 2002, 2005). However, as many researchers have noted, the density of the undergrowth in the regenerating forest vegetation on former arable lands varies greatly within the same climatic conditions and sometimes there are no trees at all (Utkin et al. 2002 Prévosto et al. 2011 Ruskule et al. 2012 Prishchepov et al. 2013 Novoselova et al. 2016 Baeva et al. 2017 Jagodziński et al. 2017 Telesnina et al. 2017 Kalinina et al. 2018 Maslov et al. 2018). Reasons of such striking differences in the undergrowth volume are not well investigated (Prévosto et al. 2011 Ruskule et al. 2012 Novoselova et al. 2016). We propose that grass fires are the most important factors defining the successional trajectory on formerly plowed fields. The lack of research on this topic may be due to the fact that it is not easy to register grass fires, especially in remote fields their traces can disappear rather quickly and it can be difficult to identify the occurrence of spring grass fires in an area after just several weeks. Another possible reason for the lack of researches on this topic may be that abandoned plowed land without grazing is a rare phenomenon all over the world, whereas it is common in Russia and Eastern Europe. Even so, such abandoned areas are large and their input in the general biogeochemical cycles and regular ecological processes over the world is significant.

On the whole, effects of fire on vegetation and the whole ecosystems have been studied in a vast number of works. Usually, fire is considered a natural and inevitable hazard having a fundamental role in sustaining biodiversity and fire-prone ecosystems (Bowman et al. 2009 Moritz et al. 2014 Smith et al. 2016 Prichard et al. 2017). Nevertheless, large gaps of knowledge exist in understanding species-specific responses to fire regimes, in comprehending the spatiotemporal effects on biota at different scales, and in explaining the interactions of fire regimes with ecosystem functions that are related to the structural components of an ecosystem, such as soil, water, atmosphere, and biota (Driscoll et al. 2010 Haslem et al. 2011 Moritz et al. 2014 Kelly et al. 2015 Prichard et al. 2017 Pellegrini et al. 2018). In the Central Russian Upland, grass fires are common, but remain practically non-investigated. Researches on the overgrowing of abandoned lands do not take “the fire aspect” into account (Heikkinen et al. 2014 Baeva et al. 2017 Telesnina et al. 2017 Kalinina et al. 2018), while for other regions, fire risks were found to be positively linked to agricultural abandonment and agricultural landscape fragmentation (Martínez et al. 2009 Zhang et al. 2010a Pausas and Fernández-Muñoz 2012 Pausas and Keeley 2014 Glagolev and Kogan 2016 Pavleychik 2016). Therefore, the objective of our study was to analyze the overgrowing on abandoned lands in the Central Russian Upland in connection with the grass fire history and to perform this study at the landscape and local scale. At the landscape scale, we analyzed spatial characteristics of the area in relation to the 30-year history of spring fire events which we reconstructed from archival Landsat satellite images. At the local scale, we identified adjacent abandoned fields which had been or had not been affected by fire and we sampled soil and vegetation there. Plant species diversity, soil quality, and relationships between soil and vegetation were assessed for the former fields and for mature forests bordering on the abandoned fields.

(A1) (A2) (A3) (A4) where is the arithmetic mean of xa and the RMSE is defined as (A5)