| { |
| "v1_Abstract": "Human population is expected to reach ~9 billion by 2050. The ensuing demands for water, food and energy would intensify land-use conflicts and exacerbate environmental impacts. Therefore we urgently need to reconcile our growing consumptive needs with environmental protection. Here, we explore the potential of a land-use optimisation strategy to increase global agricultural production on two major groups of crops: cereals and oilseeds. We implemented a spatially-explicit computer simulation model across 173 countries based on the following algorithm: on any cropland, always produce the most productive crop given all other crops currently being produced locally and the site-specific biophysical, economic and technological constraints to production. Globally, this strategy resulted in net increases in annual production of cereal and oilseed crops from 1.9 billion to 2.9 billion tons (46%), and from 427 million to 481 million tons (13%), respectively, without any change in total land area harvested for cereals or oilseeds. This thought experiment demonstrates that, in theory, more optimal use of existing farmlands could help meet future crop demands. In practice there might be cultural, social and institutional barriers that limit the full realisation of this theoretical potential. Nevertheless, these constraints have to be weighed against the consequences of not producing enough food, particularly in regions already facing food shortages.", |
| "v2_Abstract": "Human population is expected to reach 9.1 billion by 2050. The ensuing demands for water, food and energy would intensify land-use conflicts and exacerbate environmental impacts. Therefore we urgently need to reconcile our growing consumptive needs with environmental protection. Here, we explore the potential of a land-use optimisation strategy to increase global agricultural production on two major groups of crops: cereals and oilseeds. We implemented a spatially-explicit computer simulation model across 173 countries based on the following algorithm: on any cropland, always produce the most productive crop given all other crops currently being produced locally and the site-specific biophysical, economic and technological constraints to production. Globally, this strategy resulted in net increases in annual production of cereal and oilseed crops from 1.9 billion to 2.9 billion tons (46%), and from 427 million to 481 million tons (13%), respectively, without any change in total land area harvested for cereals or oilseeds. This thought experiment demonstrates that, in theory, more optimal use of existing farmlands could help meet future crop demands. In practice there might be cultural, social and institutional barriers that limit the full realisation of this theoretical potential. Nevertheless, these constraints have to be weighed against the consequences of not producing enough food, particularly in regions already facing food shortages.", |
| "v1_text": "results : Globally, our strategy resulted in net increases in annual production of cereal and oilseed crops from 1.9 billion to 2.9 billion tons (46%), and from 427 million to 481 million tons (13%), respectively, without any change in total land area harvested for cereals (651 million ha) or oilseeds (184 million ha) (Tables S1-S4). Accordingly, annual production of vegetable oil and protein meal (the primary products of oilseeds) increased from 86 million to 94 million tons (10%), and from 176 million to 228 million tons (29%), respectively. Global demand for cereals is projected to increase to 2.7 billion tons by 2030, and to 3 billion tons by 2050 (including its use as animal feed) (FAO 2006). As such, land-use optimisation could contribute substantially to meeting future demands for cereals (at least until 2030). In contrast, the modest benefits of optimisation for vegetable oil production would not be sufficient to meet expected demands in 2030 (216 million tons) or 2050 (293 million tons) (FAO 2006). 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 PeerJ reviewing PDF | (v2013:07:656:1:1:NEW 19 Sep 2013) R ev ie w in g M an us cr ip t Among cereal crops, maize and rice underwent the largest expansions in harvested area, accompanied by increases in annual crop production by 746 million tons and 560 million tons, respectively (Fig. 2a). All other cereal crops, with the exception of sorghum, declined in both area (by at least 50%) and production (by 44-53%; Figure 1a). Although the harvested area for sorghum declined by 16 million ha (40%), annual production increased by 22 million tons (38%). This is due to an increase in sorghum\u2019s average annual yield from 1.7 to 3.4 tons/ha, as a result of land-use optimisation (Fig. S1). In the case of oilseeds, soy was the only crop that expanded in area (by 60 million ha or 81%; Fig. 2b). Soy was also the only oilseed crop to experience a decrease in average annual yield (from 2.4 to 2.2 tons/ha), as most of its expansion occurred on lands that were sub-optimal for soy but still more productive under soy than under any other crop (Fig. S2). Even so, annual production of soy increased by 98 million tons (60%). Oil palm production also increased by 23 million tons (20%; Fig. 2b). We next assessed whether the benefits of land-use optimisation would be manifested where most required, by exploring its implications for cereal production in five regions of the world that face the most severe food shortages, and would likely continue to do so in the future. These regions, which include South Asia, China, Southeast Asia, East Africa and Central Africa, contain 75% (657 million) of the world\u2019s malnourished people (FAO 2012; Lobell et al. 2008). We found that in South Asia, China and Southeast Asia, rice would remain the dominant cereal crop (Fig. 3). In fact, rice-cultivated area would increase from 129 million to 176 million ha at the expense of wheat, millet and sorghum which, incidentally, are thought to be the most vulnerable to climate change impacts in South Asia (Lobell et al. 2008) (Fig. 3). In East Africa and Central Africa, maize would no longer be the dominant cereal crop. Instead, East Africa would grow mainly rice (3.1 million ha), maize (3 million ha) and barley (2.7 million ha), while Central Africa would specialise in wheat (2.2 million ha) and rice (1.1 million ha) (Fig. 3). Following land-use 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 PeerJ reviewing PDF | (v2013:07:656:1:1:NEW 19 Sep 2013) R ev ie w in g M an us cr ip t optimisation, Central Africa would almost double its annual production of cereals (Fig. 3). The other four regions would also experience increases in cereal production (11-68%) (Fig. 3). Discussion We recognise that besides productivity, other cultural and socio-political considerations also determine actual land use and production systems. For example, land-use optimisation entails reducing annual rice production in Thailand from 24.4 million to 5.9 million tons (Table S3). Rice farmers in Thailand might be hesitant to switch to planting other cereal crops, as rice has a long history of cultivation and consumption in the region, in the same way that maize is intimately associated with the cultures and history of the Americas. To explore the effects of such cultural constraints, we re-ran the model with a modified algorithm, which excluded rice-cultivated areas from the optimisation process. In this case optimised global annual production for cereals was 2.7 billion tons, slightly less than the 2.9 billion tons projected based on the optimisation of all six cereal crops. Thus the exclusion of rice from optimisation has little overall impact on the production of cereals. The specialisation of production systems implies homogenisation of farms and agricultural landscapes. Yet some farmers might prefer to maintain multiple crops for various reasons, including balancing dietary requirements, and bet-hedging against outbreaks of pests and diseases, adverse weather conditions and price fluctuations that a monoculture might be more sensitive to. While land use optimisation might indeed drive homogenisation within individual farmlands, it is not necessarily so at national and regional scales: there is considerable variation in crop diversity following optimisation, with diversity actually increasing in many countries and regions (Figs. 3-5). We do not necessarily advocate that nations should pursue, solely, a production maximization strategy, but rather our results indicate the potential for substantial increases in crop production with such an approach. In practice, other considerations, including the benefits of maintaining diverse cropping systems, will necessarily affect the agricultural decisions taken. Neither do we imply that land-use optimisation is the only solution. On the contrary, a move towards optimisation should be implemented alongside other solutions, such as closing yield gaps, which are especially high for maize in Sub-Saharan Africa (World Bank 2008). In fact, land-use optimisation needs to be complemented by improvements in farming technologies and institutional structures, such as education, and market and financial risk management systems, all of which farmers need to make best use of the land and technologies available to them. Furthermore, given that smallholder farming often is the most common form of agricultural organisation, especially (but not only) in the tropics, smallholders will need to be 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 PeerJ reviewing PDF | (v2013:07:656:1:1:NEW 19 Sep 2013) R ev ie w in g M an us cr ip t integrated in any land-use optimisation approach through the provision of education, technology, and market and finance opportunities. In conclusion, our assessment demonstrates that in theory future crop demands, at least for cereals, can be substantially met on existing agricultural land area through the pursuit of more optimal use of farmlands. In practice there might be cultural, social and institutional barriers that limit the full realisation of this theoretical potential. Nevertheless, these constraints have to be weighed against the consequences of not producing enough food, particularly in regions already facing food shortages. 150 151 152 153 154 155 156 157 PeerJ reviewing PDF | (v2013:07:656:1:1:NEW 19 Sep 2013) R ev ie w in g M an us cr ip t material and methods : We assessed geospatial information on current land-use and crop-yield for these crops at the farmland scale across 173 countries (Monfreda et al. 2008). We based our analyses on a published global geospatial dataset at 5 arc-minute resolution (~10\u00d710 km grid cell) that depicts, for the year 2000, the proportion of harvested area and actual yield reported for each crop in each grid cell (Monfreda et al. 2008). We overlaid these data to produce a new data layer of intersected polygons (i.e. land areas sharing unique geospatial information on observed yield for each crop; referred to as farmlands in text). For cereals, these data encompass a total area of 651 million ha (~42% of Earth\u2019s total arable and permanent croplands) (FAO 2012) and comprise 788,557 data polygons (polygon mean area=826\u00b14.1 ha [\u00b1 standard error]); for oilseeds, these data encompass a total area of 184 million ha and comprise 426,000 data polygons (mean area=433\u00b12.9 ha). We carried out a three-step procedure to estimate optimised crop production within each farmland (Fig. 1). First, we established a baseline of current total production of cereal or oilseed crops within each farmland. We did so by multiplying harvested area with observed yield of each cereal or oilseed crop (Monfreda et al. 2008). Unlike cereals, whereby cereal grain is the prime economically-important product, oilseeds are produced for both oil and meal. We calculated vegetable oil and protein meal production amounts by multiplying crop production with an oil- or meal-conversion factor (derived from 2008/09 data on global crop, oil and meal production) (USDA-FAS 2011). Second, we identified an optimal cereal or oilseed crop within each farmland. The optimal cereal crop was the one with the highest observed yield within each farmland. In identifying an optimal oilseed crop, we assessed relative productivity based on the combined quantity of oil and meal produced. Given that global demand for protein meal is higher than that for vegetable oil, in optimising for oilseed production, we ascribed meal a relative weightage of 1.77 tons for every ton of oil produced (derived from 2008/09 data on global crop, oil and meal consumption) (USDA-FAS 2011). 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 PeerJ reviewing PDF | (v2013:07:656:1:1:NEW 19 Sep 2013) R ev ie w in g M an us cr ip t Third, we simulated land use optimisation by converting each farmland to a monoculture of the identified optimal cereal or oilseed crop. We calculated the resultant cereal or oilseed production for each optimised farmland by multiplying total harvested area with observed yield of the optimal crop. We calculated the benefits of land use optimisation by comparing this new production volume with the baseline production prior to optimisation. In our analyses we made the following assumptions. First, we assumed that the reported yield of each crop is uniform within each farmland (mean area=826\u00b14.1 ha for cereals and 433\u00b12.9 ha for oilseeds). If parts of a farmland had substantially lower (or higher) than the reported yield for the optimal crop, we would have overestimated (or underestimated) the benefits of land use optimisation. Second, we assumed that site-specific biophysical, economic and technological constraints to production are also uniform at the scale of each farmland, such that optimisation of a farmland for any optimal crop would not be limited by, for example, variations in water scarcity or soil nutrient levels across different parts of a farmland. transformative optimisation of agricultural land use to meet future food demands : Human population is expected to reach ~9 billion by 2050. The ensuing demands for water, food and energy would intensify land-use conflicts and exacerbate environmental impacts. Therefore we urgently need to reconcile our growing consumptive needs with environmental protection. Here, we explore the potential of a land-use optimisation strategy to increase global agricultural production on two major groups of crops: cereals and oilseeds. We implemented a spatially-explicit computer simulation model across 173 countries based on the following algorithm: on any cropland, always produce the most productive crop given all other crops currently being produced locally and the site-specific biophysical, economic and technological constraints to production. Globally, this strategy resulted in net increases in annual production of cereal and oilseed crops from 1.9 billion to 2.9 billion tons (46%), and from 427 million to 481 million tons (13%), respectively, without any change in total land area harvested for cereals or oilseeds. This thought experiment demonstrates that, in theory, more optimal use of existing farmlands could help meet future crop demands. In practice there might be cultural, social and institutional barriers that limit the full realisation of this theoretical potential. Nevertheless, these constraints have to be weighed against the consequences of not producing enough food, particularly in regions already facing food shortages. PeerJ reviewing PDF | (v2013:07:656:1:1:NEW 19 Sep 2013) R ev ie w in g M an us cr ip t Lian Pin Koh1,2*, Thomas Koellner3, and Jaboury Ghazoul1 1Department of Environmental Systems Science, ETH Zurich, CHN G 73.1, Universit\u00e4tstrasse 16, Zurich 8092, Switzerland 2Current address: Woodrow Wilson School of Public and International Affairs, Princeton University, Robertson Hall, Princeton, New Jersey 08544-1013, USA 3Faculty of Biology, Chemistry and Geosciences, University of Bayreuth, Universitaetstrasse 30, 95440 Bayreuth, Germany *Corresponding author: Lian Pin Koh, Woodrow Wilson School of Public and International Affairs, Princeton University, Robertson Hall, Princeton, New Jersey 08544-1013, USA , phone: +16097590952, email: lianpinkoh@gmail.com 1 2 3 4 5 6 7 8 9 10 PeerJ reviewing PDF | (v2013:07:656:1:1:NEW 19 Sep 2013) R ev ie w in g M an us cr ip t Introduction By 2050, global human population will have grown from the current ~7 billion to ~9 billion people (United Nations 2008). These people will require more food (Evans 2009; Godfray et al. 2010). They are also likely to demand a higher proportion of meat and dairy products that require more land, water and energy to produce (Royal Society of London 2009; Tilman et al. 2001). Meeting this demand is daunting by virtue of the need to reduce greenhouse-gas emissions (Meinshausen et al. 2009), minimise fertiliser and pesticide inputs (Moss 2007), and avoid further impacts on natural ecosystems and wildlife (Ehrlich & Pringle 2008). Additionally, we might have to cope with the yet unclear implications of climate change on food security (Brown & Funk 2008; Lobell et al. 2008; Parry et al. 2004). These challenges might be met by closing yield gaps (i.e. difference between potential and actual yields) or raising yield ceilings, reducing food lost to waste, and switching to less protein-rich or more aquaculture-based diets (Foley et al. 2011; Godfray et al. 2010). Additionally, we propose that a complementary approach is to maximise agricultural returns by planting crops that are best suited to site-specific conditions. While this strategy might seem obvious, the degree to which agricultural land use is optimised and the benefits of optimisation have not been evaluated at a global scale by which benefits might be maximally realised. To test the efficacy of this land-use optimisation approach, we developed a spatiallyexplicit computer simulation model based on the following algorithm: on any cropland, always produce the most productive crop given all other crops currently being produced locally and the site-specific biophysical, economic and technological constraints to production. By evaluating crops based on their realised yields, the algorithm captures both the local biophysical limitations to production (e.g. the need for irrigation), and the behaviour of farmers in response to these constraints (e.g. the decision to irrigate or not). Therefore, for a farmer who is currently growing barley, maize, wheat and irrigated rice on his land, and if irrigated rice has the highest per-hectare realised yield given local conditions, then land-use optimisation would entail devoting the entire farmland to irrigated rice production. An implicit requirement of this approach is that goods being considered are fungible, such that individual units of different crops within a commodity group (e.g. cereals or vegetable oil) are mutually substitutable. Therefore, we illustrate our approach by optimising land use within each of two groups of essential and fungible food crops: cereals (barley, maize, millet, rice, sorghum and wheat) and oilseeds (soy, cottonseed, rapeseed, sunflower seed, groundnut and oil palm). We optimised land use by replacing all currently harvested area, for cereals or oilseeds, with the most productive crop in the set of currently harvested crops within each farmland (Fig. 1) (Monfreda et al. 2008). 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 PeerJ reviewing PDF | (v2013:07:656:1:1:NEW 19 Sep 2013) R ev ie w in g M an us cr ip t", |
| "v2_text": "results : Globally, our strategy resulted in net increases in annual production of cereal and oilseed crops from 1.9 billion to 2.9 billion tons (46%), and from 427 million to 481 million tons (13%), respectively, without any change in total land area harvested for cereals (651 million ha) or oilseeds (184 million ha) (Tables S1-S4). Accordingly, annual production of vegetable oil and protein meal (the primary products of oilseeds) increased from 86 million to 94 million tons (10%), and from 176 million to 228 million tons (29%), respectively. Global demand for cereals is projected to increase to 2.7 billion tons by 2030, and to 3 billion tons by 2050 (including its use as animal feed) (FAO, 2006). As such, land-use optimisation could contribute substantially to meeting future demands for cereals (at least until 2030). In contrast, the modest benefits of optimisation for vegetable oil production would not be sufficient to meet expected demands in 2030 (216 million tons) or 2050 (293 million tons) (FAO, 2006). Among cereal crops, maize and rice underwent the largest expansions in harvested area, accompanied by increases in annual crop production by 746 million tons and 560 million tons, respectively (Fig. 1a). All other cereal crops, with the exception of sorghum, declined in both area (by at least 50%) and production (by 44-53%; Figure 1a). Although the harvested area for sorghum declined by 16 million ha (40%), annual production increased by 22 million tons (38%). This is due to an increase in sorghum\u2019s average annual yield from 1.7 to 3.4 tons/ha, as a result of land-use optimisation (Fig. S1). PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t 5 5 In the case of oilseeds, soy was the only crop that expanded in area (by 60 million ha or 81%; Fig. 1b). Soy was also the only oilseed crop to experience a decrease in average annual yield (from 2.4 to 2.2 tons/ha), as most of its expansion occurred on lands that were sub-optimal for soy but still more productive under soy than under any other crop (Fig. S2). Even so, annual production of soy increased by 98 million tons (60%). Oil palm production also increased by 23 million tons (20%; Fig. 1b). We next assessed whether the benefits of land-use optimisation would be manifested where most required, by exploring its implications for cereal production in five regions of the world that face the most severe food shortages, and would likely continue to do so in the future. These regions, which include South Asia, China, Southeast Asia, East Africa and Central Africa, contain 75% (657 million) of the world\u2019s malnourished people (FAO, 2012, Lobell et al. , 2008 ). We found that in South Asia, China and Southeast Asia, rice would remain the dominant cereal crop (Fig. 2). In fact, rice-cultivated area would increase from 129 million to 176 million ha at the expense of wheat, millet and sorghum which, incidentally, are thought to be the most vulnerable to climate change impacts in South Asia (Lobell et al. , 2008 ) (Fig. 2). In East Africa and Central Africa, maize would no longer be the dominant cereal crop. Instead, East Africa would grow mainly rice (3.1 million ha), maize (3 million ha) and barley (2.7 million ha), while Central Africa would specialise in wheat (2.2 million ha) and rice (1.1 million ha) (Fig. 2). Following land-use optimisation, Central Africa would almost double its annual production of cereals (Fig. 2). The other four regions would also experience increases in cereal production (11-68%) (Fig. 2). Discussion We recognise that besides productivity, other cultural and socio-political considerations also determine actual land use and production systems. For example, land-use optimisation entails reducing annual rice production in Thailand from 24.4 million to 5.9 million tons (Table S3). Rice farmers in Thailand might be hesitant to switch to planting other cereal crops, as rice has a long history of cultivation and consumption in the region, in the same way that maize is intimately associated with the cultures and history of the Americas. To explore the effects of such cultural constraints, we re-ran the model with a modified algorithm, which excluded rice-cultivated areas from the optimisation process. In this case PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t 6 6 optimised global annual production for cereals was 2.7 billion tons, slightly less than the 2.9 billion tons projected based on the optimisation of all six cereal crops. Thus the exclusion of rice from optimisation has little overall impact on the production of cereals. Some might also argue against the specialisation of production systems because it implies homogenisation of agricultural landscapes. While this is true within individual farmlands, it is not necessarily so at national and regional scales: there is considerable variation in crop diversity following optimisation, with diversity actually increasing in many countries and regions (Figs. 2-4). Neither do we imply that land-use optimisation is the only solution. On the contrary, a move towards optimisation should be implemented alongside other solutions, such as closing yield gaps, which are especially high for maize in Sub-Saharan Africa (World Bank, 2008). In fact, land-use optimisation needs to be complemented by improvements in farming technologies and institutional structures, such as education, and market and financial risk management systems, all of which farmers need to make best use of the land and technologies available to them. Furthermore, given that smallholder farming often is the most common form of agricultural organisation, especially (but not only) in the tropics, smallholders will need to be integrated in any land-use optimisation approach through the provision of education, technology, and market and finance opportunities. In conclusion, our assessment demonstrates that in theory future crop demands, at least for cereals, can be substantially met on existing agricultural land area through the pursuit of more optimal use of farmlands. In practice there might be cultural, social and institutional barriers that limit the full realisation of this theoretical potential. Nevertheless, these constraints have to be weighed against the consequences of not producing enough food, particularly in regions already facing food shortages. Our results suggest that a combined scientific and socioeconomic approach that targets smallholder farmers, especially those in developing countries, might contribute to ensuring future food security and alleviating rural poverty. PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t 7 7 figure legends : Figure 2. Benefits of land-use optimisation in food-insecure regions. Left: changes in relative proportions of harvested area of cereal crops; right: changes in total crop production (numbers in parentheses indicate percentage change). Figure 3. Cereal crop diversity. Crop diversity was calculated based on the reciprocal Simpson\u2019s diversity index, \u2211\u2212 21 ip , whereby p represents the proportional area of the ith crop type. This index reflects the probability that two randomly chosen cropland areas are not cultivated for the same cereal crop. Circle areas reflect relative total harvested area for cereal crops. Dashed line indicates no change in crop diversity between current and optimal land uses. For clarity of presentation, not all country labels are shown. Figure 4. Oilseed crop diversity. Crop diversity was calculated based on the reciprocal Simpson\u2019s diversity index, \u2211\u2212 21 ip , whereby p represents the proportional area of the ith crop type. This index reflects the probability that two randomly chosen cropland areas are not cultivated for the same oilseed crop. Circle areas reflect relative total harvested area for oilseed crops. Dashed line indicates no change in crop diversity between current and optimal land uses. For clarity of presentation, not all country labels are shown. PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t Figure 1 Harvested area and crop production. Changes in cultivated area and production amounts of cereal (a) and oilseed crops (b) under current and optimal land-use allocations. Filled bars represent current land use; open bars represent optimal land use; and numbers in parentheses indicate percentage change. PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t Figure 2 Benefits of land-use optimisation in food-insecure regions. Left: changes in relative proportions of harvested area of cereal crops; right: changes in total crop production (numbers in parentheses indicate percentage change). PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t Figure 3 material and methods : We assessed geospatial information on current land-use and crop-yield for these crops at the farmland scale across 173 countries (Monfreda et al. , 2008 ). We based our analyses on a published global geospatial dataset at 5 arc-minute resolution (~10\u00d710 km grid cell) that depicts, for the year 2000, the proportion of harvested area and actual yield reported for each crop in each grid cell (Monfreda et al. , 2008 ). We overlaid these data to produce a new data layer of intersected polygons (i.e. land areas sharing the same geospatial information; referred to as farmlands in text). For cereals, these data encompass a total area of 651 million ha (~42% of Earth\u2019s total arable and permanent croplands) (FAO, 2012) and comprise 788,557 data polygons (polygon mean area=826\u00b14.1 ha [\u00b1 standard error]); for oilseeds, these data encompass a total area of 184 million ha and comprise 426,000 data polygons (mean area=433\u00b12.9 ha). To calculate current crop production within each polygon, we multiplied harvested area with observed yield (Monfreda et al. , 2008 ). In the case of oilseeds, vegetable oil and protein meal production amounts were calculated by multiplying crop production with an oil- or meal-conversion factor (derived from 2008/09 data on global crop, oil and meal production) (USDA-FAS, 2011). Unlike cereals, whereby cereal grain is the prime economically-important product, oilseeds are produced for both oil and meal. Therefore, in identifying an optimal oilseed crop, we assessed relative productivity based on the combined quantity of oil and meal produced. Given that global demand for protein meal is higher than that for vegetable oil, in optimising for oilseed production, we ascribed PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t 4 4 meal a relative weightage of 1.77 tons for every ton of oil produced (derived from 2008/09 data on global crop, oil and meal consumption) (USDA-FAS, 2011). transformative optimisation of agricultural land use to meet future food demands : Human population is expected to reach 9.1 billion by 2050. The ensuing demands for water, food and energy would intensify land-use conflicts and exacerbate environmental impacts. Therefore we urgently need to reconcile our growing consumptive needs with environmental protection. Here, we explore the potential of a land-use optimisation strategy to increase global agricultural production on two major groups of crops: cereals and oilseeds. We implemented a spatially-explicit computer simulation model across 173 countries based on the following algorithm: on any cropland, always produce the most productive crop given all other crops currently being produced locally and the site-specific biophysical, economic and technological constraints to production. Globally, this strategy resulted in net increases in annual production of cereal and oilseed crops from 1.9 billion to 2.9 billion tons (46%), and from 427 million to 481 million tons (13%), respectively, without any change in total land area harvested for cereals or oilseeds. This thought experiment demonstrates that, in theory, more optimal use of existing farmlands could help meet future crop demands. In practice there might be cultural, social and institutional barriers that limit the full realisation of this theoretical potential. Nevertheless, these constraints have to be weighed against the consequences of not producing enough food, particularly in regions already facing food shortages. PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t 1 1 Lian Pin Koh1,2*, Thomas Koellner3, and Jaboury Ghazoul1 1Department of Environmental Systems Science, ETH Zurich, CHN G 73.1, Universit\u00e4tstrasse 16, Zurich 8092, Switzerland 2Current address: Woodrow Wilson School of Publica and International Affairs, Princeton University, Robertson Hall, Princeton, New Jersey 08544-1013, USA 3Faculty of Biology, Chemistry and Geosciences, University of Bayreuth, Universitaetstrasse 30, 95440 Bayreuth, Germany *Corresponding author: Lian Pin Koh, Woodrow Wilson School of Publica and International Affairs, Princeton University, Robertson Hall, Princeton, New Jersey 08544-1013, USA , phone: +16097590952, email: lianpinkoh@gmail.com PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t 2 2 Introduction By 2050, global human population will have grown from the current 6.9 billion to 9.1 billion people (United Nations, 2008). These people will require more food (Evans, 2009, Godfray et al. , 2010 ). They are also likely to demand a higher proportion of meat and dairy products that require more land, water and energy to produce (Royal Society of London, 2009, Tilman et al. , 2001 ). Meeting this demand is daunting by virtue of the need to reduce greenhouse-gas emissions (Meinshausen et al. , 2009 ), minimise fertiliser and pesticide inputs (Moss, 2007), and avoid further impacts on natural ecosystems and wildlife (Ehrlich & Pringle, 2008). Additionally, we might have to cope with the yet unclear implications of climate change on food security (Brown & Funk, 2008, Lobell et al. , 2008 , Parry et al. , 2004 ). These challenges might be met by closing yield gaps (i.e. difference between potential and actual yields) or raising yield ceilings, reducing food lost to waste, and switching to less protein-rich or more aquaculture-based diets (Foley et al. , 2011 , Godfray et al. , 2010 ). Additionally, we propose that a complementary approach is to maximise agricultural returns by planting crops that are best suited to site-specific conditions. While this strategy might seem obvious, the degree to which agricultural land use is optimised and the benefits of optimisation have not been evaluated at a global scale by which benefits might be maximally realised. To test the efficacy of this land-use optimisation approach, we developed a spatially-explicit computer simulation model based on the following algorithm: on any cropland, always produce the most productive crop given all other crops currently being produced locally and the site-specific biophysical, economic and technological constraints to production. By evaluating crops based on their realised yields, the algorithm captures both the local biophysical limitations to production (e.g. the need for irrigation), and the behaviour of farmers in response to these constraints (e.g. the decision to irrigate or not). Therefore, for a farmer who is currently growing barley, maize, wheat and irrigated rice on his land, and if irrigated rice has the highest per-hectare realised yield given local conditions, then land-use optimisation would entail devoting the entire farmland to irrigated rice production. An implicit requirement of this approach is that goods being considered are fungible, such that individual units of different crops within a commodity group (e.g. cereals or vegetable oil) are mutually substitutable. Therefore, we illustrate our approach by optimising land use within each of two groups of essential and fungible food crops: cereals (barley, maize, millet, rice, sorghum and wheat) and oilseeds (soy, PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t 3 3 cottonseed, rapeseed, sunflower seed, groundnut and oil palm). We optimised land use by replacing all currently harvested area, for cereals or oilseeds, with the most productive crop in the set of currently harvested crops within each farmland (Monfreda et al. , 2008 ). cereal crop diversity. : Crop diversity was calculated based on the reciprocal Simpson\u2019s diversity index. This index reflects the probability that two randomly chosen cropland areas are not cultivated for the same cereal crop. Circle areas reflect relative total harvested area for cereal crops. Dashed line indicates no change in crop diversity between current and optimal land uses. For clarity of presentation, not all country labels are shown. PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t Figure 4 oilseed crop diversity. : Crop diversity was calculated based on the reciprocal Simpson\u2019s diversity index. This index reflects the probability that two randomly chosen cropland areas are not cultivated for the same oilseed crop. Circle areas reflect relative total harvested area for oilseed crops. Dashed line indicates no change in crop diversity between current and optimal land uses. For clarity of presentation, not all country labels are shown. PeerJ reviewing PDF | (v2013:07:656:0:0:NEW 17 Jul 2013) R ev ie w in g M an us cr ip t", |
| "url": "https://peerj.com/articles/189/reviews/", |
| "review_1": "Gavin Stewart \u00b7 Oct 8, 2013 \u00b7 Academic Editor\nACCEPT\nThank for you for clarifying the points raised by myself and the reviewers, and for constructive engagement with peer review. I think this paper is a great addition to the literature.", |
| "review_2": "Gavin Stewart \u00b7 Sep 5, 2013 \u00b7 Academic Editor\nMINOR REVISIONS\nThe work is a useful contribution, but responding to the reviewers comments will result in improvements. i would encourage you to amend the manuscript to address the points they make where possible particularly in the case of reviewer one. Reviewer three is critical of your methods for model selection. I would suggest that you clarify the approach you adopted in the methods and results and include some discussion of the range of sensitivities. If you wish to add more discussion of the uncertainties surrounding model choice, it might be worth discussing reversible jump MCMC as an alternative to model averaging, perhaps as a suggestion for further work. Highlighting the dangers inherent in using p values or stepwise approaches would also be useful. I do not think that you need to adopt a new strategy, simply explain the potential problems a little more. Finally I would also encourage you to run a web of science search, to make sure that no meta-analyses are available on this topic. yours truly\n\nGavin Stewart", |
| "review_3": "Reviewer 1 \u00b7 Aug 26, 2013\nBasic reporting\nThe article meets all basic reporting standards. It is well written, clear, and appropriate for publication.\nExperimental design\nI am uncertain on the methodology. I am not familiar with animal tracking or movement ecology to any real extent. However, use of rangefinder cameras or some other sampling methodology to confirm that the method used herein, tracks in show, was effective. What if animals move a lot less when it has snowed? or move very differently? This could have been easily validated as a method and would have made this paper far more useful.\nValidity of the findings\nValid findings provided the methodology effectively estimates animal movements.\nAdditional comments\nGood paper. Useful. Please consider linking to theory a bit more with respect to habitat frag, connectivity theory, circuit theory for conservation, or whatever is appropriate. Also, are there any other additional stats to be added to explore impermeability of roads or natural mitigation opportunities that may be available to us to promote movement, ie any covariates in landscape etc.\n\nPlease also add #total number of tracks per species etc. I think you just present the crossings per km. Finally, can you apply rarefaction curves or some other measure to explore accumulation of species or tracks?\nCite this review as\nAnonymous Reviewer (2013) Peer Review #1 of \"Using multi-scale distribution and movement effects along a montane highway to identify optimal crossing locations for a large-bodied mammal community (v0.1)\". PeerJ https://doi.org/10.7287/peerj.189v0.1/reviews/1", |
| "review_4": "Paul Beier \u00b7 Aug 23, 2013\nBasic reporting\nThis paper describes and illustrates methods for identifying optimal locations for highway-crossing structures for a diverse mammal community. It demonstrates the superiority of free, remotely-sensed data to costly hand-digitized data, the superiority of multi-scale models to single scale models, and presents one approach to integrating across single-species models. The methods are appropriate (with one possible exception) and clearly described. Inferences are well-grounded in the data and statistical analyses.\nThe intro states a major objective that is not mentioned in the Abstract. The abstract should state: \u201cFreely-available remotely-sensed habitat landscape data were better than more costly, manually-digitized microhabitat maps in supporting models that identified good crossing sites; however models using both types of data were better yet.\u201d The Abstract should also mention that in 6 of 8 cases the multi-scale models performed better than models at any single scale.\nThe paper is generally well-written and clear. I think you should feel free to be a bit more punchy and colloquial. For instance, instead of dryly referring to the \u201chighway and off-highway transect models,\u201d you could refer to the latter as a model of the probability that an animal approaches the highway, and the former as a model that an animal, having reached the highway, crosses it. The title should highlight the idea of identifying optimal locations for crossing structures. And the table captions should go much further in telling the reader \u201cwhat\u2019s the point.\u201d Right now every caption is \u201cHere are some numbers. Guess why we put them there. Guess what they mean.\u201d See Kroodsma (2000. Auk 117:1081-1083).\nTable 1 can be improved. State that the unit of measurement is \u201c% of area\u201d unless otherwise stated. What is the unit of measurement for road/path, and railroad? Replace the acronym column with a \u201cvariable name\u201d column that uses clear descriptions (\u201c% unvegetated\u201d instead of \u201cno_veg\u201d and \u201c% dense conifer forest\u201d instead of \u201ccon_d\u201d). Reverse the column orders (1Variable name, 2Variable description, 3Source). In the caption or in a table footnote, state that the PAP variables were hand-digitized from 1-m photos (just to remind the reader).\nTable 2. Spell out words. If \u201cInd crossings\u201d means \u201csuccessful crossings\u201d use that term (spelled out).\nTable 3, 4, 5. Captions too terse.\nTable 3 Caption should explain the point of the table.\nIn Table 5, write out the variable names; transpose rows and columns so the names will fit.\nYou seem to use \u201ccovariate\u201d and \u201cparameter\u201d and \u201cpredictor\u201d and occasionally \u201ccharacteristic\u201d interchangeably, for the same things that are more commonly referred to as \u201cexplanatory variables\u201d or \u201cvariables.\u201d Not a big deal, but I think the terms \u201cvariables\u201d and \u201cpredictors\u201d are less-offputting and more accessible to most readers. Line 167: \u201clandscape parameter\u201d should probably be \u201cland cover or condition\u201d?\nExperimental design\nMy one real concern is with averaging the maps predicting the best crossing sites for individual species to produce a map intended to serve all species. If a site A is excellent (10 on a scale of 10) for 9 species and 0/10 for the 10th species, the average is 9. If site B scores 8 for all species, its average is 8. But I\u2019d argue that B is better, no? On lines 338-9, you state that landscape traits associated with preferred crossing sites differed among species, suggesting that some species could be poorly served by a site that serves the majority. Please discuss limitations and advantages of averaging as a way of combining predictions across species, and alternative (existing or potential) ways of combining across species. Perhaps a post-hoc procedure to identify \u201closers\u201d and accommodate their needs.\nOther than this one issue, this is well-designed and well-executed. Thank you for acknowledging me for feedback on the GIS analyses, even though I do not recall providing any such feedback. At this point I feel you have more to offer me on this topic than I can offer you.\nValidity of the findings\nThe Discussion might mention that the road permeability calculation almost certainly overestimates road permeability because the denominator (# of crossings of transects 10-900m from highway, mean 175 m from highway) does not reflect the numbers of animals who avoided the road at an average distance > 175 m.\nIn the abstract, I suggest deleting \u201clikely resulting in population fragmentation\u201d unless you present evidence this is true. Puma data are too sparse to support inference about fragmentation. For other species, perhaps 10% passage rates are enough to avert fragmentation.\nAdditional comments\nThank you for the opportunity to review this fine contribution.\nCite this review as\nBeier P (2013) Peer Review #2 of \"Using multi-scale distribution and movement effects along a montane highway to identify optimal crossing locations for a large-bodied mammal community (v0.1)\". PeerJ https://doi.org/10.7287/peerj.189v0.1/reviews/2", |
| "pdf_1": "https://peerj.com/articles/189v0.2/submission", |
| "pdf_2": "https://peerj.com/articles/189v0.1/submission", |
| "review_5": "Reviewer 3 \u00b7 Aug 14, 2013\nBasic reporting\nThe subject, methods, and results for this paper are presented clearly. The paper is very well written with only a few typographical errors. Some components of the methods, in particular sample design, could be more clearly explained. I make specific comments below for the authors.\nExperimental design\nI had no concerns with the substantive components of the paper. The data and methods fit the objectives very nicely. Results were followed by a lucid discussion of the findings. My primary concern is the choice of method for model selection and validation.\n\nThe authors conduct an exhaustive analysis of the 4 types of count distribution. Interesting, but often these details are subsumed in the unreported elements of an analysis.\n\nThe hybrid information theoretic(IT)/p-value approach for identifying the most \u2018parsimonious\u2019 model was overly complex, obtuse and technically flawed. The authors selected/combined models using a combination of best subsets, screening based on the p-values of the covariates, and finally model averaging. The best subsets method can result in models that are particular to a set of data \u2013 this is useful for exploring data, but not developing models that generalise to other study areas or time periods. Choosing individual variables based on their p-values assumes that the coefficient and SE are independent of other variables in the model. This is not the case \u2013 each covariate is dependent on the contribution of other covariates. Furthermore, this mixes two model selection philosophies: IT and hypothesis testing. I don\u2019t know if a simpler model selection process (i.e., identify 15-20 models based on existing literature and theory and select best model using AICc delta) would change the results, but the current approach is certainly awkward if not statistically incorrect.\n\nAIC provides a relative not absolute measure of model fit. Thus, we have some idea of what is the best model of the set, but not if that model has any ecological or predictive validity (i.e., even the best model could be a very poor predictor). The best model should be tested against a set of withheld data; this might involve testing the observed versus predicted probability for each count or by looking at the simple residuals (observed minus predicted counts for the withheld data). This is especially important when considering that these models will be used to predict counts on the highway and guide the location of mitigation strategies.\nValidity of the findings\nAlthough the model selection process was odd and perhaps incorrect (see previous section), the sampling and data are valid and the general statistical treatment was correct. The results support the objectives and conclusions of this work.\nAdditional comments\nGeneral Comments\nThe subject, methods, and results for this paper are presented clearly. The paper is very well written with only a few typographical errors. Some components of the methods, in particular sample design, could be more clearly explained. I make specific comments below for the authors.\n\nI had no concerns with the substantive components of the paper. The data and methods fit the objectives very nicely. Results were followed by a lucid discussion of the findings. My primary concern is the choice of method for model selection and validation.\n\nThe authors conduct an exhaustive analysis of the 4 types of count distribution. Interesting, but often these details are subsumed in the unreported elements of an analysis.\n\nThe hybrid information theoretic/p-value approach for identifying the most \u2018parsimonious\u2019 model was overly complex, obtuse and technically flawed. The authors selected/combined models using a combination of best subsets, screening based on the p-values of the covariates, and finally model averaging. The best subsets method can result in models that are particular to a set of data \u2013 this is useful for exploring data, but not developing models that generalise to other study areas or time periods. Choosing individual variables based on their p-values assumes that the coefficient and SE are independent of other variables in the model. This is not the case \u2013 each covariate is dependent on the contribution of other covariates. Furthermore, this mixes two model selection philosophies: IT and hypothesis testing. I don\u2019t know if a simpler model selection process (i.e., identify 15-20 models based on existing literature and theory and select best model using AICc delta) would change the results, but the current approach is certainly awkward if not statistically incorrect.\n\nAIC provides a relative not absolute measure of model fit. Thus, we have some idea of what is the best model of the set, but not if that model has any ecological or predictive validity (i.e., even the best model could be a very poor predictor). The best model should be tested against a set of withheld data; this might involve testing the observed versus predicted probability for each count or by looking at the simple residuals (observed minus predicted counts for the withheld data). This is especially important when considering that these models will be used to predict counts on the highway and guide the location of mitigation strategies.\n\nDetailed Comments\nAbstract\nThe abstract is well written and provides a balanced summary of the key objectives and results of the paper. However, some additional description of results (1 additional sentence) would be useful.\nL6: comma should follow \u201cCanada\u201d\n\nIntroduction\nL57-59: This sentence is somewhat confusing and should be simplified \u2013 doesn\u2019t modelling animal movement result in predictive models?\nL60: Unclear why there is an emphasis on migratory birds and reptiles; there is much published evidence of multi-scale habitat selection and movement by mammals as well.\nL77-84: The authors provide a very nice statement of the study objectives.\n\nMethods\nL104: The number and type of sampling locations should be a simple idea, but it is not clearly explained. In particular, why were transects surveyed versus the edge of a road and how do these 10 transects relate to the 9 transects on L128? Why is there a reference to a right-of-way \u2013 is this the highway or another type of linear feature? Did the authors survey highways and off-road/gravel roads?\nL108-109: The methods for sampling tracks/transects is key to understanding the results and should be reported not referenced.\nL202-208: Not necessary to derive the Vuong test; a citation will suffice.\nL217-231: This is an overly complex approach for model selection that some would classify as data dredging. Using a classic IT approach, each model should serve as a hypothesis. The approach used by the authors might result in a very predictive model, but the best subsets approach will likely capitalise on unique correlations that limit the generalisation of the \u2018best model\u2019. Furthermore, the method of selecting variables based on individual p-values is unusual and fails to consider the covariation of the covariates (i.e., the magnitude and variance of any one covariate is a function of other covariates in the model, thus, they cannot be considered as unique entities).\nL237: AIC provides a relative not absolute measure of model fit. See comment above.\nL239: What is meant by \u201cexhaustive model tests\u201d?\nL248: The identity/source of the polygon is unclear; is a polygon a pixel? If so, then just refer to pixels.\nL253: Are the \u201cmammal group estimates\u201d, estimates of the counts of crossings?\n\nResults\nL265-266: I recommend presenting tables in parentheses only; no reason to waste text to introduce a table.\nL279: Did the authors screen for excessively high multicollinearity using VIF/tolerance scores? The inclusion of the same variable at the three scales suggests that this might be a problem.\n\nDiscussion\nThe discussion was well formulated. However, there is much reference to mitigation strategies, but no mention of what those strategies might be. Two or three sentences describing mitigation would be of value to those readers not familiar with highway planning.\n\nTables and Figures\nTable 5: Not clear what is meant by \u201cselection frequencies\u201d; are these the averaged coefficients from the best models?\nThe figure captions are too brief \u2013 they should state where and what, relative to the objectives of the study.\n\nFigure 2: how does one differentiate between the predictive scores for the transect and the highway?\nCite this review as\nAnonymous Reviewer (2013) Peer Review #3 of \"Using multi-scale distribution and movement effects along a montane highway to identify optimal crossing locations for a large-bodied mammal community (v0.1)\". PeerJ https://doi.org/10.7287/peerj.189v0.1/reviews/3", |
| "all_reviews": "Review 1: Gavin Stewart \u00b7 Oct 8, 2013 \u00b7 Academic Editor\nACCEPT\nThank for you for clarifying the points raised by myself and the reviewers, and for constructive engagement with peer review. I think this paper is a great addition to the literature.\nReview 2: Gavin Stewart \u00b7 Sep 5, 2013 \u00b7 Academic Editor\nMINOR REVISIONS\nThe work is a useful contribution, but responding to the reviewers comments will result in improvements. i would encourage you to amend the manuscript to address the points they make where possible particularly in the case of reviewer one. Reviewer three is critical of your methods for model selection. I would suggest that you clarify the approach you adopted in the methods and results and include some discussion of the range of sensitivities. If you wish to add more discussion of the uncertainties surrounding model choice, it might be worth discussing reversible jump MCMC as an alternative to model averaging, perhaps as a suggestion for further work. Highlighting the dangers inherent in using p values or stepwise approaches would also be useful. I do not think that you need to adopt a new strategy, simply explain the potential problems a little more. Finally I would also encourage you to run a web of science search, to make sure that no meta-analyses are available on this topic. yours truly\n\nGavin Stewart\nReview 3: Reviewer 1 \u00b7 Aug 26, 2013\nBasic reporting\nThe article meets all basic reporting standards. It is well written, clear, and appropriate for publication.\nExperimental design\nI am uncertain on the methodology. I am not familiar with animal tracking or movement ecology to any real extent. However, use of rangefinder cameras or some other sampling methodology to confirm that the method used herein, tracks in show, was effective. What if animals move a lot less when it has snowed? or move very differently? This could have been easily validated as a method and would have made this paper far more useful.\nValidity of the findings\nValid findings provided the methodology effectively estimates animal movements.\nAdditional comments\nGood paper. Useful. Please consider linking to theory a bit more with respect to habitat frag, connectivity theory, circuit theory for conservation, or whatever is appropriate. Also, are there any other additional stats to be added to explore impermeability of roads or natural mitigation opportunities that may be available to us to promote movement, ie any covariates in landscape etc.\n\nPlease also add #total number of tracks per species etc. I think you just present the crossings per km. Finally, can you apply rarefaction curves or some other measure to explore accumulation of species or tracks?\nCite this review as\nAnonymous Reviewer (2013) Peer Review #1 of \"Using multi-scale distribution and movement effects along a montane highway to identify optimal crossing locations for a large-bodied mammal community (v0.1)\". PeerJ https://doi.org/10.7287/peerj.189v0.1/reviews/1\nReview 4: Paul Beier \u00b7 Aug 23, 2013\nBasic reporting\nThis paper describes and illustrates methods for identifying optimal locations for highway-crossing structures for a diverse mammal community. It demonstrates the superiority of free, remotely-sensed data to costly hand-digitized data, the superiority of multi-scale models to single scale models, and presents one approach to integrating across single-species models. The methods are appropriate (with one possible exception) and clearly described. Inferences are well-grounded in the data and statistical analyses.\nThe intro states a major objective that is not mentioned in the Abstract. The abstract should state: \u201cFreely-available remotely-sensed habitat landscape data were better than more costly, manually-digitized microhabitat maps in supporting models that identified good crossing sites; however models using both types of data were better yet.\u201d The Abstract should also mention that in 6 of 8 cases the multi-scale models performed better than models at any single scale.\nThe paper is generally well-written and clear. I think you should feel free to be a bit more punchy and colloquial. For instance, instead of dryly referring to the \u201chighway and off-highway transect models,\u201d you could refer to the latter as a model of the probability that an animal approaches the highway, and the former as a model that an animal, having reached the highway, crosses it. The title should highlight the idea of identifying optimal locations for crossing structures. And the table captions should go much further in telling the reader \u201cwhat\u2019s the point.\u201d Right now every caption is \u201cHere are some numbers. Guess why we put them there. Guess what they mean.\u201d See Kroodsma (2000. Auk 117:1081-1083).\nTable 1 can be improved. State that the unit of measurement is \u201c% of area\u201d unless otherwise stated. What is the unit of measurement for road/path, and railroad? Replace the acronym column with a \u201cvariable name\u201d column that uses clear descriptions (\u201c% unvegetated\u201d instead of \u201cno_veg\u201d and \u201c% dense conifer forest\u201d instead of \u201ccon_d\u201d). Reverse the column orders (1Variable name, 2Variable description, 3Source). In the caption or in a table footnote, state that the PAP variables were hand-digitized from 1-m photos (just to remind the reader).\nTable 2. Spell out words. If \u201cInd crossings\u201d means \u201csuccessful crossings\u201d use that term (spelled out).\nTable 3, 4, 5. Captions too terse.\nTable 3 Caption should explain the point of the table.\nIn Table 5, write out the variable names; transpose rows and columns so the names will fit.\nYou seem to use \u201ccovariate\u201d and \u201cparameter\u201d and \u201cpredictor\u201d and occasionally \u201ccharacteristic\u201d interchangeably, for the same things that are more commonly referred to as \u201cexplanatory variables\u201d or \u201cvariables.\u201d Not a big deal, but I think the terms \u201cvariables\u201d and \u201cpredictors\u201d are less-offputting and more accessible to most readers. Line 167: \u201clandscape parameter\u201d should probably be \u201cland cover or condition\u201d?\nExperimental design\nMy one real concern is with averaging the maps predicting the best crossing sites for individual species to produce a map intended to serve all species. If a site A is excellent (10 on a scale of 10) for 9 species and 0/10 for the 10th species, the average is 9. If site B scores 8 for all species, its average is 8. But I\u2019d argue that B is better, no? On lines 338-9, you state that landscape traits associated with preferred crossing sites differed among species, suggesting that some species could be poorly served by a site that serves the majority. Please discuss limitations and advantages of averaging as a way of combining predictions across species, and alternative (existing or potential) ways of combining across species. Perhaps a post-hoc procedure to identify \u201closers\u201d and accommodate their needs.\nOther than this one issue, this is well-designed and well-executed. Thank you for acknowledging me for feedback on the GIS analyses, even though I do not recall providing any such feedback. At this point I feel you have more to offer me on this topic than I can offer you.\nValidity of the findings\nThe Discussion might mention that the road permeability calculation almost certainly overestimates road permeability because the denominator (# of crossings of transects 10-900m from highway, mean 175 m from highway) does not reflect the numbers of animals who avoided the road at an average distance > 175 m.\nIn the abstract, I suggest deleting \u201clikely resulting in population fragmentation\u201d unless you present evidence this is true. Puma data are too sparse to support inference about fragmentation. For other species, perhaps 10% passage rates are enough to avert fragmentation.\nAdditional comments\nThank you for the opportunity to review this fine contribution.\nCite this review as\nBeier P (2013) Peer Review #2 of \"Using multi-scale distribution and movement effects along a montane highway to identify optimal crossing locations for a large-bodied mammal community (v0.1)\". PeerJ https://doi.org/10.7287/peerj.189v0.1/reviews/2\nReview 5: Reviewer 3 \u00b7 Aug 14, 2013\nBasic reporting\nThe subject, methods, and results for this paper are presented clearly. The paper is very well written with only a few typographical errors. Some components of the methods, in particular sample design, could be more clearly explained. I make specific comments below for the authors.\nExperimental design\nI had no concerns with the substantive components of the paper. The data and methods fit the objectives very nicely. Results were followed by a lucid discussion of the findings. My primary concern is the choice of method for model selection and validation.\n\nThe authors conduct an exhaustive analysis of the 4 types of count distribution. Interesting, but often these details are subsumed in the unreported elements of an analysis.\n\nThe hybrid information theoretic(IT)/p-value approach for identifying the most \u2018parsimonious\u2019 model was overly complex, obtuse and technically flawed. The authors selected/combined models using a combination of best subsets, screening based on the p-values of the covariates, and finally model averaging. The best subsets method can result in models that are particular to a set of data \u2013 this is useful for exploring data, but not developing models that generalise to other study areas or time periods. Choosing individual variables based on their p-values assumes that the coefficient and SE are independent of other variables in the model. This is not the case \u2013 each covariate is dependent on the contribution of other covariates. Furthermore, this mixes two model selection philosophies: IT and hypothesis testing. I don\u2019t know if a simpler model selection process (i.e., identify 15-20 models based on existing literature and theory and select best model using AICc delta) would change the results, but the current approach is certainly awkward if not statistically incorrect.\n\nAIC provides a relative not absolute measure of model fit. Thus, we have some idea of what is the best model of the set, but not if that model has any ecological or predictive validity (i.e., even the best model could be a very poor predictor). The best model should be tested against a set of withheld data; this might involve testing the observed versus predicted probability for each count or by looking at the simple residuals (observed minus predicted counts for the withheld data). This is especially important when considering that these models will be used to predict counts on the highway and guide the location of mitigation strategies.\nValidity of the findings\nAlthough the model selection process was odd and perhaps incorrect (see previous section), the sampling and data are valid and the general statistical treatment was correct. The results support the objectives and conclusions of this work.\nAdditional comments\nGeneral Comments\nThe subject, methods, and results for this paper are presented clearly. The paper is very well written with only a few typographical errors. Some components of the methods, in particular sample design, could be more clearly explained. I make specific comments below for the authors.\n\nI had no concerns with the substantive components of the paper. The data and methods fit the objectives very nicely. Results were followed by a lucid discussion of the findings. My primary concern is the choice of method for model selection and validation.\n\nThe authors conduct an exhaustive analysis of the 4 types of count distribution. Interesting, but often these details are subsumed in the unreported elements of an analysis.\n\nThe hybrid information theoretic/p-value approach for identifying the most \u2018parsimonious\u2019 model was overly complex, obtuse and technically flawed. The authors selected/combined models using a combination of best subsets, screening based on the p-values of the covariates, and finally model averaging. The best subsets method can result in models that are particular to a set of data \u2013 this is useful for exploring data, but not developing models that generalise to other study areas or time periods. Choosing individual variables based on their p-values assumes that the coefficient and SE are independent of other variables in the model. This is not the case \u2013 each covariate is dependent on the contribution of other covariates. Furthermore, this mixes two model selection philosophies: IT and hypothesis testing. I don\u2019t know if a simpler model selection process (i.e., identify 15-20 models based on existing literature and theory and select best model using AICc delta) would change the results, but the current approach is certainly awkward if not statistically incorrect.\n\nAIC provides a relative not absolute measure of model fit. Thus, we have some idea of what is the best model of the set, but not if that model has any ecological or predictive validity (i.e., even the best model could be a very poor predictor). The best model should be tested against a set of withheld data; this might involve testing the observed versus predicted probability for each count or by looking at the simple residuals (observed minus predicted counts for the withheld data). This is especially important when considering that these models will be used to predict counts on the highway and guide the location of mitigation strategies.\n\nDetailed Comments\nAbstract\nThe abstract is well written and provides a balanced summary of the key objectives and results of the paper. However, some additional description of results (1 additional sentence) would be useful.\nL6: comma should follow \u201cCanada\u201d\n\nIntroduction\nL57-59: This sentence is somewhat confusing and should be simplified \u2013 doesn\u2019t modelling animal movement result in predictive models?\nL60: Unclear why there is an emphasis on migratory birds and reptiles; there is much published evidence of multi-scale habitat selection and movement by mammals as well.\nL77-84: The authors provide a very nice statement of the study objectives.\n\nMethods\nL104: The number and type of sampling locations should be a simple idea, but it is not clearly explained. In particular, why were transects surveyed versus the edge of a road and how do these 10 transects relate to the 9 transects on L128? Why is there a reference to a right-of-way \u2013 is this the highway or another type of linear feature? Did the authors survey highways and off-road/gravel roads?\nL108-109: The methods for sampling tracks/transects is key to understanding the results and should be reported not referenced.\nL202-208: Not necessary to derive the Vuong test; a citation will suffice.\nL217-231: This is an overly complex approach for model selection that some would classify as data dredging. Using a classic IT approach, each model should serve as a hypothesis. The approach used by the authors might result in a very predictive model, but the best subsets approach will likely capitalise on unique correlations that limit the generalisation of the \u2018best model\u2019. Furthermore, the method of selecting variables based on individual p-values is unusual and fails to consider the covariation of the covariates (i.e., the magnitude and variance of any one covariate is a function of other covariates in the model, thus, they cannot be considered as unique entities).\nL237: AIC provides a relative not absolute measure of model fit. See comment above.\nL239: What is meant by \u201cexhaustive model tests\u201d?\nL248: The identity/source of the polygon is unclear; is a polygon a pixel? If so, then just refer to pixels.\nL253: Are the \u201cmammal group estimates\u201d, estimates of the counts of crossings?\n\nResults\nL265-266: I recommend presenting tables in parentheses only; no reason to waste text to introduce a table.\nL279: Did the authors screen for excessively high multicollinearity using VIF/tolerance scores? The inclusion of the same variable at the three scales suggests that this might be a problem.\n\nDiscussion\nThe discussion was well formulated. However, there is much reference to mitigation strategies, but no mention of what those strategies might be. Two or three sentences describing mitigation would be of value to those readers not familiar with highway planning.\n\nTables and Figures\nTable 5: Not clear what is meant by \u201cselection frequencies\u201d; are these the averaged coefficients from the best models?\nThe figure captions are too brief \u2013 they should state where and what, relative to the objectives of the study.\n\nFigure 2: how does one differentiate between the predictive scores for the transect and the highway?\nCite this review as\nAnonymous Reviewer (2013) Peer Review #3 of \"Using multi-scale distribution and movement effects along a montane highway to identify optimal crossing locations for a large-bodied mammal community (v0.1)\". PeerJ https://doi.org/10.7287/peerj.189v0.1/reviews/3\nReview 6: \n", |
| "LLM_extracted_review": "[\"Review 1: The paper is a great addition to the literature.\\n\\nReview 2: The work is a useful contribution, but responding to the reviewers' comments will result in improvements. Clarify the approach for model selection and include discussion of uncertainties surrounding model choice. Highlight dangers of using p-values or stepwise approaches. Run a web of science search for existing meta-analyses on the topic.\\n\\nReview 3: The article meets basic reporting standards and is well written. Methodology is uncertain; validation of tracking methods is needed. Consider linking to habitat fragmentation and connectivity theory. Additional statistics on road impermeability and natural mitigation opportunities should be included. Present total number of tracks per species and apply rarefaction curves for species accumulation.\\n\\nReview 4: The paper demonstrates the superiority of free, remotely-sensed data and multi-scale models. The introduction states a major objective not mentioned in the abstract. Improve table captions and clarify variable names. Discuss limitations of averaging predictions across species. The road permeability calculation may overestimate actual permeability. Delete unsupported claims about population fragmentation.\\n\\nReview 5: The paper is well written but some methods could be clearer. The model selection process is overly complex and potentially flawed. A simpler model selection process is recommended. AIC provides a relative measure of model fit; the best model should be tested against withheld data. Additional description in the abstract would be useful. Clarify sampling methods and the relationship between transects. The discussion should include specific mitigation strategies.\\n\\nReview 6: [No content provided for Review 6.]\"]" |
| } |
