Representing large-scale land acquisitions in land use change scenarios for the Lao PDR

Agricultural large-scale land acquisition (LSLA) is a process that is currently not captured by land change models. We present a novel land change modeling approach that includes processes governing LSLAs and simulates their interactions with other land systems. LSLAs differ from other land change processes in two ways: (1) their changes affect hundreds to thousands of contiguous hectares at a time, far surpassing other land change processes, e.g., smallholder agriculture, and (2) as policy makers value LSLA as desirable or undesirable, their agency significantly affects LSLA occurrence. To represent these characteristics in a land change model, we allocate LSLAs as multi-cell patches to represent them at scale while preserving detail in the representation of other dynamics. Moreover, LSLA land systems are characterized to respond to an explicit political demand for LSLA effects, in addition to a demand for various agricultural commodities. The model is applied to simulate land change in Laos until 2030, using three contrasting scenarios: (1) a target to quadruple the area of LSLA, (2) a moratorium for new LSLA, and (3) no target for LSLA. Scenarios yield drastically different land change trajectories despite having similar demands for agricultural commodities. A high level of LSLA impedes smallholders’ engagement with rubber or cash crops, while a moratorium on LSLA results in increased smallholder involvement in cash cropping and rubber production. This model goes beyond existing land change models by capturing the heterogeneity of scales of land change processes and the competition between different land users instigated by LSLA. Electronic supplementary material The online version of this article (10.1007/s10113-018-1316-8) contains supplementary material, which is available to authorized users.

Changing demands for land system goods, services, and effects are resolved by CLUMondo by allocating or removing land systems that can provide these demands, as defined in the land system definition and classification (see this supplement, part 2). It does this in an iterative procedure where the land system with the highest transition potential at that time and location is allocated. The transition potential of land systems producing demands that are not met are increased and vice versa, after which the procedure is repeated in a next iteration. Transition potential adjustments continue until all demands are met within a 10% error margin, and the average deviation of all demands is lower than 5%.
Transition potential is determined by the sum of three land system specific factors: (1) location suitability, (2) conversion resistance, and (3) competitive advantage. We describe these factors consecutively below.
Location suitability is quantified using logistic regressions, performed in this study using R software. 28 explanatory biophysical and socioeconomic factors have been collected from various data sources (see this supplement, part 3). These factors are given in two versions: a normal version and a smoothed version used to assess suitability for large-scale systems. Location suitability quantifies how suitable a certain location is for a certain land system. For each land system, a set of explanatory factors that relate significantly to the current location of that land system is selected. For example, in this study, we found that the current location of small rubber plantations are located on location accessible to cities and towns, with higher precipitation, and a low hazard of river flooding (see this supplement, table S4). By using logistic regression models, suitability can be quantified between 0 and 1 using these significant factors.
Conversion resistance quantifies the resistance a land system has to being converted to another land system and is related to, among others, capital investment. For example, urban land systems are typically highly resistant to change, while forest systems are more easily converted. Conversion resistance is a value between 0 and 1 and is quantified based on expert judgement (see this supplement, table S7).
Competitive advantage is initially zero for all land systems, but changes upwards or downwards during iterations. When the demands that a certain land system provides are not met, the parameter is adjusted up, and when there is overproduction of these demands, it is adjusted down. This continues until all demands are met within the defined margin of error (10% in this application and 5% on average over all demands). Note that CLUMondo does not rely on a predefined hierarchy to handle competition between different demands (one demand is not more important than another).
CLUMondo is furter constrained by the conversion matrix. Not all changes are allowed (e.g. water is not allowed to change to anythin else), and some changes take a minimum number of years to be allowed (e.g. reforestation takes a number of years). These settings are given in this supplement, table S5.
Conversion order ranks for every land system how competitive it is to deliver a specific demand. When a demand is not met, the competitive advantage of the system with the highest rank for that demand is increased more than the competitive advantge of lower-ranking systems. This ensures logical trajectories of change. Land systems can also have a negative rank (-1) for a specific demand, which means that they are not considered to resolve a deviation in demand, even though they may produce that demand (for example, in this application, small plantation systems also produce subsistence crops because they are, at the scale of the 400ha cell, in a mosaic with smallholder farmers. However, an increase in demand for subsistence crops should not result in the allocation of more small plantations). The conversion order is given in this supplement, table S6.

Large-scale land acquisitions data pre-processing
LSLA data was collected from the Land Observatory (www.landobservatory.org) and the Centre for Development and the Environment (CDE, Switzerland). In order to avoid errors of commission (inclusion of LSLAs that do not exist) we cross-checked the two input datasets, and consulted Google Earth satellite imagery in case of doubt. Reported LSLAs that lacked spatial accuracy (e.g. only province reported) were dropped. 396 LSLAs were thus identified. For 306 data entries, only point data was available. In this case, we created buffers of the reported area around the centroids to approach the shape. For 90 data entries, the boundary of the LSLA was available as polygon data, as part of an on-going survey (Hett 2015).

Smallholder agriculture classifications
In the land system classification decision tree, a cash crop focused smallholder system is defined as a cell where the fraction of the area that is covered by cash crops is larger than 25%. For this purpose, the sum of area fractions as reported in the Lao Agricultural Census 2010 for the following crops is calculated. - Similarly, smallholder areas where rubber covers over 25% of the area are classified as rubberpermanent smallholder mosaic. For swidden, we reclassified the land system map by Ornetsmüller, Verburg, and Heinimann (2016). This map represents a number of swidden systems with differing intensities and forest cover, which we reclassified to a single swidden system.

Land system service calculations
For all 15 land systems, the provision of each of five land system services (timber, rubber, cash crops, subsistence crops and urban area) was quantified. These services remain constant in all scenarios, whereas the LSLA service is scenario-dependent. Exception to this is the timber service in the Moratorium scenario which is kept constant, because smallholders cannot substitute as a supplier in our application.
In a first step, area breakdowns of land covers per land system were empirically established. In the cases of urban, water, dense forest and bare land, it was assumed that these land systems are covered 100% by their respective land covers. The same assumption was used for all large plantation systems. For small plantation systems (including coffee plantations), overlay analysis using the actual polygons of the plantation and the plantation land system cells was performed to determine the average area percentage of LSLAs within LSLA land system cells. Similarly, the area dedicated to cash crops by smallholders was quantified by overlaying the agricultural census (GoL, 2011) with the land system raster. The same operation was used to calculate average tree cover of land systems, using the national land cover map (GoL 2010). Subsistence crops were calculated as the remainder area for each land system.
For LSLA in Laos, it is known that the granted area (the polygons used in this study) is often much larger than the allocated area, which is again larger than the developed area. An inventory for two Lao provinces indicates that current LSLAs use only 49% (Luang Prabang) and 12% (Xiengkhouang) of their granted area (Hett 2015). Expansion beyond the granted area also occurs but is much more rare. As there are no nationwide statistics, we quantified productive use for LSLAs by overlaying a forest map (GoL, 2010) with arable plantation polygons. This way we established that, for small arable plantations, on average 18% of the granted area is covered with trees and therefore not used productively. For large arable plantations, 45% is similarly not used productively. As the same method cannot be used for rubber, coffee or timber plantations, as the land cover map records these land uses as tree cover, we used the same number for all small and large plantations respectively, where coffee plantations are considered small plantations.
Next, typical yields were used calculate the average services output per land system. These yields are assumed to be constant across all cells belonging to the same system.