Enhancing the ecological value of tropical agriculture through 4 set-asides 5

Agricultural expansion across the tropics is the primary driver of biodiversity declines and 27 ecosystem service degradation. However, efforts to mitigate these negative impacts may reduce 28 commodity production. We quantify trade-offs between oil palm cultivation and ecological 29 outcomes (biodiversity, above-ground carbon storage and dung nutrient cycling) across 30 different potential set-aside (uncultivated areas in agricultural landscapes) strategies. We show 31 that all set-aside configurations yield substantial gains in ecological outcomes. The best 32 strategy involves spatially targeted riparian reserves, such as those used in oil palm certification 33 schemes, where species occurrence can be doubled without reducing overall cultivation area. 34 Adopting this strategy throughout the 8 million hectares of plantations in Borneo would lead 35 to extensive improvements in ecological outcomes without losses to production area, and 36 consequently, enhancing agricultural sustainability. 37

. Before 113 evaluating the impacts of set-aside policies, we consulted major producers across the palm oil 114 industry to ensure our analyses focused on maximum slopes for cultivation and riparian reserve 115 widths that could be implemented feasibly in a real-world context (Methods). We then 116 examined different set-aside configurations, encompassing combinations of 20 riparian reserve 117 widths (in 5 m increments, ranging from 5 to 100 m, on both sides of rivers) and 11 maximum 118 slope angles (ranging from 15 to 25°) per plantation, equating to 880 combinations across the 119 four plantations (220 in each). Larger riparian reserve widths and lower maximum slopes for 120 cultivation both mean that there is a greater area of set-aside in the landscape and, 121 correspondingly, less area available for cultivating crops. 122 123 Across all the set-aside configurations examined, 61-92% of the landscape remained available 124 for cultivation. We find that riparian reserve set-aside comprised 0.5 to 10% of the landscape, 125 while set-aside based on maximum slope for cultivation accounted for 4 to 30%. By 126 comparison, 20 and 50 m riparian reserve widths (corresponding, respectively, to current 127 policies for Sabah in Malaysia and Indonesia), combined with 25° maximum slope, would 128 leave 89-91% of the landscape available for oil palm cultivation ( Fig. 2A  To assess the trade-off between the land available for cultivation, and ecological outcomes 133 (biodiversity, ecosystem function and ecosystem service), we combined our set-aside 134 configurations with field-derived distributions for 235 species (150 birds, 19 non-volant 135 mammals, 21 bats and 45 dung beetles), dung nutrient cycling and LiDAR-derived above-136 ground forest carbon storage (Methods). We express ecological outcomes of different 7 landscape scenarios in terms of net and relative percentage changes in species occurrence, and 138 total above-ground forest carbon storage and dung nutrient cycling under different set-aside 139 configurations. Net changes in species occurrence are calculated as the percentage change in 140 landscape area, whereas relative changes are calculated as a percentage change in species area. 141 For example, if a species occurred in 20% of the landscape in one set-aside configuration and 142 then 30% in another, this would equate to a 10% net increase and a 33% relative increase. 143

144
We evaluated two categories of set-aside policies. First, we considered 'uniform' policy 145 scenarios, meaning a one-size-fits-all approach, as implemented in most national/state-scale 146 legislation. Even with these very simple policies, the potential importance of set-asides in 147 delivering ecological gains in tropical agricultural landscapes becomes clear. In our landscape, 148 each 10% of the area in set-aside results in a net increase in species occurrence, ranging from 149 3% to 23% across all 235 species (mean = 10% net increase, but up to 223% relative increase), 150 a 6% net increase in above-ground carbon storage, and 9% net increase in dung nutrient cycling 151 ( Fig. 3 and Fig. 4A blue curve). We also evaluated 'variable' policies, under which we 152 optimized set-aside configurations, allowing these to vary between plantations in a way that 153 maximized ecological outcomes at least cost to cultivation. To calculate the variable policy 154 outcomes, we used multi-objective optimization models to maximize ecological outcomes 155 (species occurrence, above-ground carbon storage and dung nutrient cycling across set-aside 156 in the landscape) (objective one) and maximize area of the landscape available for oil palm 157 cultivation (objective two). 158 159 Compared to a uniform approach, the variable policy yields even higher levels of species 160 occurrence and above-ground carbon storage for any given percentage of the landscape 161 cultivated. Alternatively, the variable policy achieves specified levels of species occurrence 162 8 and above-ground carbon storage at lower overall set-aside area than in the equivalent uniform 163 policy (Fig. 4). The greatest gains from the variable policy are obtained when set-aside 164 configurations result in 77-87% of the landscape cultivated (upper quartile of the difference 165 between uniform and variable policies; Fig. 4A,B). The most efficient of these is achieved 166 when 83% of the landscape is cultivated ('maximum efficient'). In this scenario, net species 167 occurrence within set-asides rises by 8.1% (range: 0.3-18% net increase in occurrence across 168 all species), from an average across species of 55% for the uniform policy to an average of 169 63% for the variable policy, and 3.8% more above-ground carbon stored (Supplementary Table  170 3; Fig. 4A,E). By comparison, achieving the same gain in ecological outcomes with the uniform 171 policy would require a reduction in cultivation area of 7.7% (Supplementary Table 3 Supplementary Note 2), with the greatest average gains among the birds, including endemic 175 and threatened species. We also find that at 90% ('business-as-usual'; broadly equivalent to 176 current policies in Indonesia and Malaysia) and 70% ('high level set-aside') of the landscape 177 cultivated ( Fig. 4C-F), the variable policy enhances ecological outcomes, albeit to a lesser 178 degree than when 83% of the landscape is planted (Fig. 4D However, the flexibility of the variable policy, allows for more spatially targeted set-asides to 185 be distributed heterogeneously across the landscape to maximize ecological outcomes. As a 186 result, the variable policy could have lower overall set-aside with a mean riparian reserve width 187 9 of 44 m and mean maximum slope for cultivation of 22° to achieve the same ecological 188 outcome (Supplementary Table 4). This is particularly pertinent for the variable set-aside 189 configurations used in most certification schemes (Supplementary Note 1), because they should 190 translate into improved ecological outcomes without the need to reduce cultivation area. 191

192
We also conducted our optimizations with a uniform maximum slope of 25° but letting riparian 193 reserve width vary. We did this because the palm oil industry told us that varying maximum 194 slopes for cultivation would be less favorable from an operational perspective. Again, the 195 variable policy resulted in ecological gains, albeit with reduced benefit compared to when the 196 maximum slope for cultivation could also vary ( Across the tropics, agricultural plantations are less likely to occur on steep slopes, because they 207 are more expensive to deforest and harder to cultivate successfully (6, 44). Landscape 208 topography is, therefore, a key attribute affecting the impacts of set-aside policies. In our study 209 landscape, each plantation had a distinct topographic profile (Supplementary Table 2 (Table 1). 227 228 Our findings are important for both conservation and food security debates as we show that 229 set-asides can greatly enhance ecological outcomes without compromising the area of the 230 landscape available for cultivation. This is critical because perceived losses to production may 231 disincentivize growers from adopting best practice set-aside measures. To this end, our study 232 shows that locally tailored riparian set-asides may be the best way to boost the biodiversity and 233 ecosystem service value of tropical agricultural landscapes.  Table 2). 252 253 Across the study area, we sampled multiple taxonomic groups, above-ground carbon storage 254 and dung nutrient cycling. Methods, locations, and sample sizes varied, but all encompassed 255 logged forest and riparian forest fragments and oil palm (details for each group or function are 256 provided below). Species occurrence data from the logged forest reserve was used to improve 257 our estimates of species distributions, but were not used in the trade-off analyses. 258

259
We obtained plantation boundaries for the experimental landscape directly from plantation 260 owners. We mapped rivers across the landscape using a combination of geographic information 261 system (GIS) data from the Sabah DID and the Shuttle Radar Topography Mission (STRM) 12 (http://srtm.usgs.gov) digital elevation model (DEM) at a resolution of 30  30 m. The DID 263 data included the location of rivers, but did not include hydrological information such as flow, 264 which is used to estimate channel width. To estimate flow, we first used the r.watershed module 265 in GRASS GIS to create raster files for flow accumulation and drainage direction, which were 266 then inputted into the r.stream.extract module to create a raster and vector of channels using 267 the flow accumulation and direction layers. We subsequently added network information to the 268 raw vector channels using an R script to find links between channels 269 (https://www.safeproject.net/dokuwiki/safe_gis/stream_networks). The STRM generated data 270 matched very closely with the governmental DID data, so we used the STRM generated river 271 network in our analysis, which allowed us to exclude small streams estimated to be under 5 m 272 in channel width, because in all guidelines and legislation these size rivers receive no or very 273 small riparian reserves. We ground-truthed 20 rivers to ensure that predictions of channel width 274 were broadly accurate. To estimate and map slope across the landscape, the SRTM data was 275 further processed using the gdaldem_slope function (https://gdal.org/programs/gdaldem.html) 276 for Python to generate a raster of slope angles measured in degrees. 277 278

Palm oil producer consultations 279
Before undertaking our landscape analyses, we consulted palm oil producers to inform the 280 range of set-asides policies to be tested, to ensure that the policies tested were feasible to 281 implement from an industry perspective. We conducted semi-structured interviews with nine 282 representatives from seven of the largest palm oil producers, with plantations located in nine 283 different countries across Southeast Asia and West Africa. Collectively, these companies 284 manage about 9% of the world's industrial palm oil plantations, an area of land covering 1.7 285 million ha. From these consultations, two key set-aside components emerged, riparian reserve 286 widths and maximum slope for cultivation. Eight of the nine respondents felt that increasing 287 13 riparian reserve width was both feasible and important for enhancing ecological outcomes 288 (biodiversity, ecosystem functions and ecosystem services). Additionally, all respondents 289 indicated that they would support the establishment of wildlife corridors within plantations, 290 with riparian reserves being the main way to achieve this. Four out of the nine respondents 291 were supportive of policy changes to maximum slope for cultivation, but explained that they 292 rarely cultivate slopes steeper than 20°. 293 294

Set-aside configurations used in the analyses 295
Set-aside configurations of maximum slopes for cultivation and riparian reserve widths were 296 assessed in a GIS. We created 20 different riparian reserve width layers by adding buffers of 297 5-100 m (in 5 m increments) around the river network. We created polygons for 11 different 298 thresholds for maximum planting slope ranging from 15 to 25° (in 1° increments). These two 299 sets of layers were subsequently merged to produce 220 combined riparian reserve width and 300 maximum slope for cultivation layers and then clipped to each plantations (but not the forest 301 reserve) to produce 880 plantation-specific set-aside layers. Across the four plantations, this 302 resulted in 220 4 or 2,342,560,000 unique ways to configure the landscape. The landscape 303 configurations were overlaid with species distributions, above-ground carbon storage and dung 304 nutrient cycling layers. These allowed us to examine and optimize trade-offs between between 305 the amount of land available for cultivation and the ecological outcomes. 306 307 Each five-meter increase in riparian reserve width results in an increase in set-aside of just 0.44 308 -0.52% of total production area, staying more-or-less constant across the 20 riparian reserve 309 widths we tested ( Fig. 2A). On the other hand, decreasing the maximum slope for cultivation 310 reduces planted area to a much greater extent, with a one-degree change leading to a 0.9-4.1% 311 reduction in cultivated area (Fig. 2B). absence prediction was then made using the sensitivity-specificity (SES) equality metric. We 354 did not use bioclimatic variables as predictors because we were working at a fine-resolution 355 landscape-scale and there was not enough variability. Instead, we used location and landcover 356 predictors (elevation, slope, distance to river and soil type), which are static and do not change 357 with the configuration of the experimental landscape. As such, our estimated species 358 distributions represent the largest possible predicted distribution for each species across the 359 landscape. Relative variable importance was computed using Pearson's correlations between 360 16 predictions of the full model and with each variable iteratively removed. All SDMs were 361 constructed using the SSDM package for R (https://www.r-project.org/). 362 363

Dung nutrient cycling predictions 364
Dung removal is an important part of the soil nutrient cycling process and reduces greenhouse 365 gas emissions (52). We measured nutrient cycling via dung removal at 309 sample points across 366 the landscape. At each location, 700 g of dung were placed under a rain cover and, 24 hours 367 later, any remaining dung was collected and weighed. We also used three 368 evaporation/precipitation controls, comprising 700 g piles which were not accessible to fauna. 369 For further details see (51). To estimate dung removal across the entire landscape, we used 370 residual corrected ordinary regression kriging between our point estimates, and landscape level 371 predictors implemented in SAGA GIS. We predicted dung removal using the same predictors 372 as for the species distribution models, plus dung beetle diversity and non-volant mammal 373 diversity (summed from our species distribution models), due to the relationship between 374 mammals and dung beetles (53,54).

Optimization of trade-offs 404
We formulated a mixed integer linear programing (MILP) model to optimize set-aside policies 405 for riparian reserve width and maximum slope for cultivation across the oil palm plantations. 406 The objective of the model is to maximize ecological outcomes in set-aside, subject to a limit 407 on the area of land taken out of cultivation and put into set-aside. The model was run for a 408 range of different set-asides to produce Pareto-optimal curves of ecological outcomes, where: . .
Model (S1)-(S4) is a modified version of what is known in the site selection literature as a 421 "maximum covering" problem (57). The objective (S1) maximizes the weighted proportional 422 ecological outcome within set-asides. Constraint (S2) sets an upper limit (aka budget) on total 423 set-aside area across the landscape. Parameter is a user-specified value that can be adjusted 424 up/down to increase/decrease the set-aside area budget. Equalities (S3) require selection of 425 exactly one policy for riparian reserve width and maximum slope for cultivation for each 426 plantation . Inequalities (S4), meanwhile, determine the fraction of each ecological outcome 427 20 within set-aside areas. Given the structure of the optimization model, constraints (S4) could be 428 written as equalities, since each variable will automatically equal the value on the right-429 hand-side. Finally, constraints (S5) impose binary restrictions on the ℓ variables for selecting 430 riparian reserve widths and maximum slopes for cultivation. 431

432
To impose a uniform policy for riparian reserve width and maximum slope for cultivation 433 across all plantations, we introduce variable ℓ equal to one if riparian reserve width and 434 maximum slope for cultivation ℓ is selected as a standard, zero otherwise, and the following 435 side constraints: 436 Equality (S6) requires selection of a uniform policy for riparian reserve width and maximum 438 slope for cultivation, while equalities (S7) stipulate that all plantations must adopt the same 439 policy. 440

441
We implemented our landscape set-aside optimizations in the OPL modeling language using 442 CPLEX studio version 12.9 (58), which employs branch-and-cut methods to solve MILPs. The 443 largest problem instance we solved had 237 continuous variables, 880 binary variables, and 444 243 constraints. We performed secondary optimization runs assuming a uniform maximum 445 slope for cultivation of 25°. We also ran a set of optimizations of specific combinations of 446 riparian reserve width and maximum slope for cultivation to test existing policies in Malaysia 447 and Indonesia. We then plotted where these lie on top of the Pareto-optimal curves. 448 449 Estimating improvements to palm oil cultivation across Borneo 450 To calculate the area of Borneo suitable for oil palm cultivation, we clipped the dataset of 451 global oil palm suitability created by (44) to Borneo and then extracted and summed the area 452 of 'Suitable', 'High', and 'Perfect' categories across the island. We then revised this figure by 453 removing existing protected areas (from https://protectedplanet.net/) and existing oil palm 454 plantations (from https://atlas.cifor.org/). We then intersected the remaining area with all areas 455 falling between 15-25° slopes (at a 90 m resolution), by following the same procedure 456 described above for assessing slopes across the study landscape. We estimated the area of 457 Borneo within 100 m of a perennial river using river networks created by Milieux 458