Long-term assessment of ecosystem services at ecological restoration sites using Landsat time series

Reversing ecological degradation through restoration activities is a key societal challenge of the upcoming decade. However, lack of evidence on the effectiveness of restoration interventions leads to inconsistent, delayed, or poorly informed statements of success, hindering the wise allocation of resources, representing a missed opportunity to learn from previous experiences. This study contributes to a better understanding of spatial and temporal dynamics of ecosystem services at ecological restoration sites. We developed a method using Landsat satellite images combined with a Before-After-Control-Impact (BACI) design, and applied this to an arid rural landscape, the Baviaanskloof in South Africa. Since 1990, various restoration projects have been implemented to halt and reverse degradation. We applied the BACI approach at pixel-level comparing the conditions of each intervened pixel (impact) with 20 similar control pixels. By evaluating the conditions before and after the restoration intervention, we assessed the effectiveness of long-term restoration interventions distinguishing their impact from environmental temporal changes. The BACI approach was implemented with Landsat images that cover a 30-year period at a spatial resolution of 30 meter. We evaluated the impact of three interventions (revegetation, livestock exclusion, and the combination of both) on three ecosystem services; forage provision, erosion prevention, and presence of iconic vegetation. We also evaluated whether terrain characteristics could partially explain the variation in impact of interventions. The resulting maps showed spatial patterns of positive and negative effects of interventions on ecosystem services. Intervention effectiveness differed across vegetation conditions, terrain aspect, and soil parent material. Our method allows for spatially explicit quantification of the long-term restoration impact on ecosystem service supply, and for the detailed visualization of impact across an area. This pixel-level analysis is specifically suited for heterogeneous landscapes, where restoration impact not only varies between but also within restoration sites.


24
Abstract: 25 Reversing ecological degradation through restoration activities is a key societal challenge of the 26 upcoming decade. However, lack of evidence on the effectiveness of restoration interventions leads to 27 inconsistent, delayed, or poorly informed statements of success, hindering the wise allocation of 28 resources, representing a missed opportunity to learn from previous experiences. This study contributes 29 to a better understanding of spatial and temporal dynamics of ecosystem services at ecological 30 restoration sites. We developed a method using Landsat satellite images combined with a Before-After- 31 Control-Impact (BACI) design, and applied this to an arid rural landscape, the Baviaanskloof in South 32 Africa. Since 1990, various restoration projects have been implemented to halt and reverse degradation. 33 We applied the BACI approach at pixel-level comparing the conditions of each intervened pixel (impact) 34 with 20 similar control pixels. By evaluating the conditions before and after the intervention, we 35 assessed the effectiveness of long-term restoration interventions distinguishing their impact from 36 environmental temporal changes. The BACI approach was implemented with Landsat images that 37 cover a 30-year period at a spatial resolution of 30 m. We evaluated the impact of three interventions 38 (revegetation, livestock exclusion, and the combination of both) on three ecosystem services; forage 39 provision, erosion prevention, and presence of iconic vegetation. We also evaluated whether terrain 40 characteristics could partially explain the variation in impact of interventions. The resulting maps 41 showed spatial patterns of positive and negative effects of interventions on ecosystem services.

51
Rural landscapes depend on and simultaneously supply several ecosystem services, nature's 52 contribution to people [1,2]. However, the quality of rural landscapes is deteriorating due to the 53 expansion of croplands and grasslands into native vegetation and unsustainable agricultural practices 54 [3]. Land degradation affects 40% of the agricultural land on earth, reducing the provision of ecosystem 55 services and resulting in adverse environmental, social, and economic consequences [4][5][6]. It has been 56 estimated that land degradation has a detrimental effect on 3.2 billion individuals and reflects an 57 economic loss in the range of 10 percent of annual global gross product [3]. Given the increased pressure 58 on ecosystems, restoration of degraded lands has become an important element of multiple global 59 initiatives [7]. Several international initiatives have developed strategic targets as part of land 60 sustainability agendas [e.g. 8-13] that are directly or indirectly linked to ecological restoration [14]. More 61 recently, the United Nations (UN) declared 2021-2030 as the Decade on Ecological Restoration, with the 62 aim of recognizing the need to accelerate global restoration of degraded ecosystems to mitigate negative 63 impacts of climate change crisis protect biodiversity on the planet [15]. However, ineffective restoration 64 efforts could inadvertently lead to a major waste of resources, continued deterioration of biodiversity 65 and perceptions of conservation failure [16]. 66 Restoration is defined as 'any intentional activity that initiates or accelerates the recovery of an 67 ecosystem from a degraded state'; regardless of the form or intensity of degradation [17]. Restoration 68 actions can vary from improving vegetation cover [e.g. 18,19] to diverse land management and policy 69 implementations for improving the quality of terrestrial [e.g. 7,20,21], aerial [e.g. 22], or aquatic 70 ecosystems [e.g. 23,24]. A successful ecological restoration should be effective, efficient and engaging 71 through a collaboration of multiple stakeholders across sectors [25]. However, the basis of evidence to 72 guide restoration practitioners is scarce, given the lack of long-term monitoring to determine the 73 circumstances under which restoration efforts work [26]. This lack of impact evidence leads to 74 incomplete, overdue or poorly informed claims of progress, hinders the effective allocation of resources 4 75 and represents a lost opportunity to select the best technologies and methods based on a critical 76 evaluation of the lessons learned [27][28][29]. 77 Monitoring and evaluating restoration interventions presents several challenges, including: i) 78 restoration projects are often implemented across large areas with limited accessibility and large spatial 79 heterogeneity; ii) the high economic costs and capacity constraints of field monitoring methodologies 80 hinder the long-term documentation of restoration projects, particularly to assess the effects of such 81 interventions outside their project timespan; iii) restoration initiatives often take a long time to start 82 generating benefits [30]; iv) simple comparison of means between impact and control sites do not 83 account for pre-existing differences between sites; v) after a restoration effort, ecosystem services show 84 great variation in their temporal and spatial patterns and rate of change of the trajectories towards the 85 desired reference [31]; and vi) and vi) observed state changes may also be attributable to intra-and inter- Despite widespread awareness of the potential of RS, most ecosystem service studies use static 99 land use/land cover information rather than a more dynamic assessment of satellite records 100 [35,37,45,46]. However, land use/land cover maps are not always available [47], can lead to 5 101 generalization errors as they exclude spatial variation within the same vegetation category [35,48], and 102 may be outdated or available only at large temporal intervals. In addition, most studies fail to take 103 advantage of the long temporal records of available remotely sensed data, one of their great strengths 104 in assessing ecosystem services [33,35]. Previous studies have shown that directly linking in situ 105 observations of ecosystem services to remotely sensed data improves the capturing of their spatio-106 temporal dynamics as compared to the often-used practice of linking the service supply directly to one 107 land cover class [49,50]. 108 A method for assessing the impacts of natural or human-induced disturbances on ecosystems 109 where the allocation of treatment and control sites cannot be randomized, is the before-after-control-110 impact/treatment (BACI) approach [51]. Among other applications, the BACI approach can be used to 111 assess the impacts of long-term restoration interventions independently of natural temporal changes 112 [52]. It compares the conditions of a restored area (impact) with the conditions of nearby unrestored 113 (control) areas before and after the restoration intervention [53,54]. The BACI approach was recently 114 applied using RS images to assess land restoration interventions in semi-arid landscapes in West-Africa 115 [32] using 20 automatically selected control sites for each impact site and multiple years for their 116 "before" and "after" periods. The use of several controls invalidates claims that the findings of the BACI 117 assessment are primarily due to a weak choice of control sites [55]. In the respective study for West 118 Africa [33], topographic variations were not explicitly accounted for, and intervention effectiveness was 119 assessed for the entire impact site without considering terrain variation and within-site differences in 120 interventions' effectiveness. These can however be important because different vegetation types grow 121 in locations with different elevation, slope, aspect, and parent material (geological material from which 122 soils are formed) [56][57][58][59][60]. The freely accessible collection of historical Landsat imagery can mitigate the 123 widespread lack of timely, long-term, reliable, and homogeneous ground information for monitoring 124 restoration interventions. 125 This study contributes to a better understanding of the spatial and long-term distribution of 126 ecosystem service supply for supporting the site-specific evaluation of restoration interventions by 6 127 expanding the spatial scope of the BACI analyses to pixel-level. By analyzing intervention impact at 30 128 m pixel-level rather than for large intervention areas, we aim to capture variation and patterns of 129 intervention outcomes within a heterogeneous landscape. The specific aims of this research are to: (1) 130 quantify the effect of restoration interventions on ecosystem service supply using Landsat time series 131 data and the BACI approach at pixel-level, and (2) evaluate whether terrain characteristics affect the 132 spatial distribution of restoration effectiveness, using an arid agricultural landscape study area in South 133 Africa as a case study. Each of the restoration interventions aimed to address local environmental challenges associated 176 with land degradation by improving ecosystem service supply. To illustrate, for this paper we selected 177 three ecosystem services; one provisioning, one regulating, and one cultural (Table 1).

232
The RS and other spatial data used in this study are summarized in Table 3 (Table S1) based on how similar the trajectory of the spectral index was between pixels. 250 We used up to 50 iterations and a convergence limit of 1. We arbitrarily limited the prior vegetation 251 characteristics to five vegetation clusters with the intention of distinguishing the key groups with 252 varying temporal behavior before the interventions took place, following the procedure described in 253 [32]. 254 To calculate the BACI contrast for the approximately 22600 intervened pixels we selected Landsat 255 images representing the greenest moment of the year. This moment is defined by calculating the average 256 highest vegetation index -or lowest bare soil index value-for the study area (i.e. maximum MSAVI or 257 minimum BSI value) (Table S2). Pixels falling within the Landsat 7 Enhanced Thematic Mapper Plus 258 (ETM+) Scan Line Corrector (SLC) off data were excluded. We considered three years for the period 259 before and after intervention, with exception of the interventions 'livestock exclusion' and 'livestock 260 exclusion + revegetation' for which not enough images were available before 1989 and consequently we 261 used only two years for the period 'before'. To focus the comparison on sites with a similar reference 262 state, the BACI analyzes was carried out separately for each cluster. Secondly, for each of the intervened 263 pixels, we obtained the spectral index values for each of the assessed years. We randomly selected 20 264 control pixels per intervened pixel [32]. We developed a simple Windows command line application to 265 randomly select the control pixels from the same vegetation clusters as the intervened site, avoiding 266 pixels within the SLC off data from Landsat 7 ETM+. Using the same command line application, we 267 extracted the spectral index values for each intervened pixel and its respective control pixels for each 268 year of the period before and after the intervention. We then calculated the BACI contrast based on the 269 following formula: 270 BACI contrast = (μ CA -μ CB ) -(μ IA -μ IB ) 271 where μ is the temporal (selected years) and spatial (20 controls) mean of the variables selected to 272 represent the impact (in this study the selected vegetation indices); the letters C and I stand for control 13 273 and impact, respectively; and the letters B and A stand for the periods "before and "after", respectively. 274 By convention, a negative contrast indicates that the variable (except the BSI index, which is a proxy for 275 percentage bare soil instead of vegetation) has increased more (or decreased less for BSI) in the impact 276 site with respect to the control sites during the time period ranging from before to after the 277 implementation of the restoration project. The BACI contrast is expressed in the same units of the 278 variable of interest, i.e. the spectral index used, and consequently is unitless in our case. 279 Since our data did not pass the Shapiro-Wilk test for normality, we used a nonparametric  Wallis test (Games-Howell post-hoc test at 0.05) to explore the differences between restoration 281 interventions, vegetation clusters, terrain aspect, and soil parent material on the BACI contrast. We 282 randomly sampled pixels for each compared group (i.e. intervention, cluster number, aspect or parent 283 material, applying a minimum distance of 60 meters between points to avoid selecting neighboring 284 pixels and ensure independent samples. We selected the five parent materials classes that represent the 285 largest areas in intervened sites (Table 4). We also checked for association between the BACI contrast of 286 each restoration intervention and ecosystem service with slope and elevation by fitting regression 287 models, using random samples of 10% of the data.

293
The RS-based models that best describe ecosystem services as measured in the field are presented 294 in Table 5. The best model for erosion prevention is a second-degree polynomial fit with the BSI index. 295 Forage provision was also fitted to second-degree polynomials using the NBR. The presence of iconic 296 species is the only ecosystem service described with a linear regression model that uses two predicting 297 variables (MSAVI and elevation), where elevation contributed by 22% to the model (expressed as partial 298 R 2 ). The R 2 in Table 5 represents the mean R 2 obtained from the repeated cross-validation. We used 299 another image matching the fieldwork period to check if the models were consistent for different dates. 300 The models based on spectral information for 17/07/2017 resulted in lower R 2 as compared to when 301 spectral information for the validation data14/05/2017 was used for erosion prevention and forage 302 provision. The R 2 of presence of iconic species increased by 2% for the validation date. 303

Comparison of BACI contrast between interventions, vegetation clusters and terrain
327 variables 328 In We also evaluated the relation between the state of the vegetation before interventions took place 349 (as indicated by their cluster) and the BACI contrast. Vegetation clusters 1 and 2 comprised areas with 350 relatively more vegetation, whereas vegetation clusters 4 and 5 contain more bare soil. No significant 351 differences between clusters were found for 'revegetation' sites for erosion prevention; and better BACI 352 contrast in vegetation clusters 4 and/or 5 for forage provision and presence of iconic species (Table 7). 353 We excluded cluster 5 to compare 'revegetation' sites for the presence of iconic species because the area  Table 7 we can observe the general tendency that 364 especially areas which had little vegetation before the restoration intervention (Cluster 4-5), show a 365 small, but significant increase in ecosystem services. 366 Regarding the assessed terrain variables to explain the intervention impact we found that south-

407
The BACI approach allows for relative comparisons of spatial and temporal differences that can be 408 used to extract the unbiased restoration impact [55,95]. However, correct understanding of the 409 underlying calculation process is needed to accurately interpret results. First, for each restoration 410 intervention, BACI was applied using a different number of years for the pre-intervention period (Table   411   S1). This will affect the BACI contrast, given that the spectral indices values vary between years a p-value that shows the significance of the BACI contrast using 20 control sites (e.g. Figure S8). We 421 compared 20 control pixels for each intervened pixel, and it is important to consider that those non-422 significant BACI contrasts can become significant when increasing the number of controls to, for 423 example,100 pixels. However, this calculation would be more time and computational consuming and 424 our explorative analysis with 100 control sites resulted in similar interventions impacts.

425
The pixel-level implementation of BACI using RS data assists in better spatially explicit 426 evaluations of restoration interventions. Our method is particularly efficient for collecting historical 427 data and evaluating large, remote, and heterogeneous areas where data collection is difficult and 428 resource consuming. Our study area is located in a dry area with relatively little cloud cover. The 429 availability of RS images will decrease in areas with frequent cloudy days, such as the in humid tropics. 430 Depending on the availability of satellite images, the 'before' and 'after' reference periods could be 431 changed or extended, allowing, for example, for the exclusion of abnormally dry years, or adding 432 another after period to evaluate the differences at different intervals of times from the start of the 433 intervention. 434 Although the levels of the BACI contrast are small when expressed as absolute numbers (e.g. forage 435 provision ranging from 0.04 to -0.11 in Table 6, translating to -9 to 11 kg per m 2 ) they can represent high 436 relative change in ecosystem supply. The relative BACI contrast (as presented in Figure S7) highlights 437 the magnitude of the contrast in relation to the analyzed pixel's spectral index value before the 438 intervention took place. The percentages in the map of Figure S7 are particularly high when the baseline 439 spectral value was close to zero. In contrast, areas that were originally more densely vegetated and still show high vegetation growth, 455 showed negative effects of the interventions. The better BACI contrasts of less vegetated areas could be 456 explained by considering that when the baseline supply of an ecosystem service is so small (e.g. 457 stratified vegetation cover of 1%), any improvement change will reflect as a great difference. In contrast, The inclusion of aspect and parent material allowed capturing differences and gaining insights for