Research, part of a special feature on Land Acquisitions, Well-being, and Environment

INTRODUCTION

Renewed interest in international land investment for agricultural use was sparked by the global financial crisis when food prices drastically increased between 2006 and 2008. It has since remained a key issue because of the increased demand for biofuels and the need to feed growing populations (Interdonato et al. 2020). Much of this interest has been focused on land in the global south, where international investors pursue crop production abroad to secure resources for domestic food and energy demand. Similarly, the host countries view large-scale land acquisitions (LSLAs) as a means to improve their own economic security while also enhancing their socio-economic development (Thiombiano et al. 2017). Many host countries depend on the agricultural sector for a high proportion of their gross domestic product (GDP) and livelihoods, which makes such investments appealing opportunities given the need for capital and modernization of agricultural practices (Schoneveld 2017). The estimates of the total land area leased since the year 2000 range from approximately 100 million to over 200 million hectares across more than 90 host countries; however, some regions have emerged as LSLA hotspots (Anseeuw et al. 2012, De Maria 2019). The majority of LSLA has occurred in the global south, particularly across much of Sub-Saharan Africa where land tenure is precarious, allowing governments to control land rights (Deininger 2011).

Land acquisitions represent a uniquely large-scale change of land tenure and land use to intensive agricultural production, thereby changing the composition and structure of the landscape and human–environment interactions (Debonne et al. 2018). The large land cover changes (LCC) associated with LSLA can exacerbate environmental issues, such as deforestation, soil degradation, climate change, and biodiversity loss, due to habitat loss (Bustamante et al. 2014, Davis et al. 2015, Shete et al. 2016). Specifically, concerns have been raised over the conversion of previously uncultivated land that was deemed available cropland for commercialization, including forests, grasslands, and protected areas (Balehegn 2015, Nalepa et al. 2017). Lazarus (2014) contended that the effects of rapid transformations on the landscape caused by LSLA could be similar to those previously exhibited by frontier clearing, which left long-term changes to the Earth’s surface and highlighted the scientific consensus surrounding the importance of understanding long-term outcomes from anthropogenic activities on land. In this vein, LCC remains a key aspect in determining the effects of LSLA given the implications for both environmental and social wellbeing.

Despite the growing awareness of the need to identify LCC trends resulting from LSLA, few studies have done so on a broad scale (Messerli et al. 2014). A significant hindrance to completing these studies stems from the lack of transparency regarding foreign investments (Interdonato et al. 2020), specifically, insufficient spatial data on precisely where acquisitions occur. Although there have been efforts made by national and international organizations to document LSLA, there is still minimal available information regarding the deals. As noted by Oya (2013), the accuracy of databases reporting LSLA must be evaluated based on their source of information, which can consist of media outlets and crowdsourced data. Some databases provide ancillary information such as the transaction date and intended size; however, the process of converting the land for agricultural use is frequently done incrementally. Moreover, the status of transactions often changes regarding the extent granted, allocated, and cultivated (Messerli et al. 2014). These problems in data acquisition and reporting have led to inconsistent and incomplete records of LSLA (Liao et al. 2020). Consequently, studies cannot be based solely on documentation.

By 2020, the Land Matrix (LM) database had documented 700 concluded or intended deals across Africa and an additional 110 that have failed. These deals total 33.9 million hectares, much of them concentrated in east Africa. Ethiopia has been a primary target country, with the highest percentage of total land grabbed among sub-Saharan Africa countries, based on Land Portal and World Bank data (Balehegn 2015). This is in part due to the government’s strong promotion of foreign investments through its inclusion in the five-year Growth and Transformation Plan II, put forth in 2016 (Federal Democratic Republic of Ethiopia 2016). The Ministry of Agriculture presented these deals as a pathway to increase food security, create jobs, and promote technology transfer, among other benefits (Keeley et al. 2014). It is not yet possible to understand the scope of socio-environmental outcomes from this foreign investment in Ethiopia. Without proper monitoring of land cover, it remains unknown how much land has been converted and the pace at which LCC is occurring.

Remotely sensed data present an opportunity to determine the extent of LCC and have been used in studies to map LCC associated with land transactions in regions across the globe (Debonne et al. 2018, Liao et al. 2020, Lay et al. 2021). Our study used national level classifications from remotely sensed data to create and analyze LCC, resulting in a representation of the pre and post land transaction landscape and an LCC map for Ethiopia. The classifications were created for 2006 and 2017, capturing the period in which a large proportion of LSLA were implemented in Ethiopia; the occurrence of land deals slowed after the year 2015 (Lay et al. 2021). The principal objectives of this study were, first, mapping the locations of LCC at a national level and, second, examining LCC at documented LSLA sites and their surrounding areas. Although similar studies have been conducted (Shete et al. 2016, Williams et al. 2020), this work differs in the scale at which land cover is classified, enabling analysis to an extent that encapsulates a larger number of LSLAs. Although the second objective focused on geographically specific LSLAs in western Ethiopia, the study more broadly provided an evaluation of current LSLA databases as a possible analysis tool and inherently addressed the advantages and disadvantages of such large-scale studies. Results provide valuable information for policymakers and researchers by estimating the area and types of LCC during the height of the land rush in Ethiopia.

STUDY AREA

Ethiopia is characterized by diverse environmental attributes that vary across the country, including climate, vegetation, and topography. Given its location near the tropics, much of the climate is controlled by the shift in the Inter-Tropical Convergence Zone and associated rain belt, which creates regionally specific precipitation regimes (Seleshi and Zanke 2004). Additional impact is generated by regional topography, which partially controls spatial trends of vegetation cover by influencing temperature as well as orographic effects. The highlands are located on either side of the rift valley, which reach altitudes between 1500 and greater than 3800 meters above sea level (ASL) and have the most favorable climate. The drier lowlands range from less than 500 to 1500 meters ASL and occur in the southeast, far west, and the northeast—notably, the Danakil Desert in the northeast, where the elevation is more than 135 meters below sea level (Viste and Sorteberg 2013). Climatic and topographic variation results in much of the cultivation occurring in the more productive western highlands, where abundant rainfall and more moderate temperatures allow for temperate mixed crop rainfed farming. Forests include highland perennial farming of various tree crops and are primarily limited to the southwest because of historically high rates of deforestation for agricultural expansion (Dixon et al. 2001, Gebru 2016).

The majority of the LSLAs in Ethiopia, as much as 80%, are concentrated in the western lowlands of the country, disregarding the less productive agroecological conditions of the region (Fig. 1). Ethiopia's LSLAs are predominately located in the states of Benishangul-Gumuz, Gambella, and lowland parts of the Southern Nations, Nationalities, and People's Region (SNNPR). These areas include many locations where intensive cultivation has not previously taken place and are at risk of frontier clearing (Keeley et al. 2014). As of 2021, the LM had recorded 151 transactions in Ethiopia, 10% of which failed and 9% that were still pending. Approximately 1.9 million hectares are registered as transaction areas and, of that, nearly 33% is domestic investment. The locations of these deals are documented in the LM database. However, for 90% of these deals, only administrative region is indicated. Exact location is reported in only 2%. Examining records from both the LM and the Ethiopian Land Bank, provided by the Ministry of Agriculture, revealed that the majority of the LSLAs were documented between 2006 and 2017. Of the 1.6 million hectares designated as currently under contract, 79% of the deals exhibited a change in land ownership after 2006, although it is important to note there is often a lengthy temporal lag between tenure change and implementation of new agricultural land use, meaning that LCC is not expected immediately after the land is leased by investors.

MATERIALS AND METHODS

Land cover classification

National level land cover maps were created for two dates that best captured the change expected before and after the height of land transactions in Ethiopia, with 2006 representing a pre-transaction state and 2017 representing post-transaction conditions. The maps were created based on a classification scheme with six classes, which included dense forest (vegetation dominated by tree cover of 70% or greater); savanna (grassland, shrubs, herbaceous plants, and tree cover less than 70%); cultivated (land used for agriculture); bare soil (bare soil and rock with < 5% vegetation cover); urban/built (land predominantly covered by man-made structures, roads, and buildings); and water/wetland (permanent water bodies, rivers, lakes, or wetland). The Google Earth Engine (GEE) cloud-computing platform was used for data processing and classification. Several explanatory variable sets based on different remote sensing sources were utilized to construct the classification models. Table 1 shows the lists of the explanatory variables of the classifications.

Explanatory variables of the 2017 classification

Landsat 8 Collection 1 surface reflectance (bands one to seven) were used as the main source of remote sensing data. The Landsat surface reflectance products are the Landsat images corrected for atmospheric effects. Time series of Landsat 8 surface reflectance images from 1 March 2016–1 March 2018 were used for time series compositing. For all images, the cloudy pixels of each image were masked using the Landsat 8 quality band, generated using the f-mask algorithm (Zhu and Woodcock 2012). A common problem when working with time series remote sensing images and images from different scenes is the radiometric inconsistencies between images. Conversion of the raw digital numbers to surface reflectance values alleviates radiometric inconsistency issues to some extent. However, some degree of inconsistency could still exist even in surface reflectance products. Khatami et al. (2020) showed that replacing the original image bands with normalized difference indices (NDI) could improve the classification accuracy of large areas. Thus, the seven spectral bands of the Landsat 8 images were replaced by six NDIs calculated by Eq. 1 using consecutive bands, as follows:

(1)

Here, i is 1 to 6. Two composites were created using the cloud-masked NDIs’ time series, including probability density function (PDF) and sequential composites. For the PDF composite, five layers, corresponding to 5%, 25%, 50%, 75%, and 95% of per-pixel values, were created. The PDF layers were created for each of the six NDIs, resulting in a total of 30 layers. For the sequential composite, first, the time series images were divided into four seasons: (1) December, January, and February; (2) March, April, and May; (3) June, July, and August; and (4) September, October, and November. Then, per-pixel median values were calculated per season for each NDI. This resulted in a sequential composite with 24 layers. In addition to the PDF and sequential composites, a per-pixel sinusoidal function was fitted to the time series of the normalized difference vegetation index (NDVI) from the Landsat 8 images. The sinusoidal function was used to model the annual greenness variations and use such variation parameters to improve the classification. Eq. 2 shows the function used for NDVI modeling.

(2)

Here, t is time, with year as the unit. Landsat 8 surface reflectance data for one year, from 1 July 2016–1 July 2017, were used for sinusoidal model fitting. The sinusoidal composite with four layers was created based on the four coefficients (α₀–α₄) from the per-pixel fitted models. Finally, the panchromatic band was used to extract textural information from Landsat images. Landsat 8 top of atmosphere images from 1 January 2017–17 January 2017 were used for textural information calculation. The GEE built-in functions were used for texture calculation, which resulted in 18 textural layers based on gray-level co-occurrence matrix (GLCM), such as entropy, variance, contrast, and homogeneity.

Landsat imagery has a 16-day revisit cycle, which can result in missing finer temporal variations. In practice, this gap can extend beyond 16 days for pixels obscured by clouds or cloud shadows during image acquisition. To reduce the number of pixels with missing data, time-series composites were generated seasonally, rather than at finer temporal resolutions, for the sequential composite and using only five percentiles for the PDF composite. To obtain denser time-series composites, daily data from the Moderate Resolution Imaging Spectroradiometer (MODIS) were also used to create both PDF and sequential composites. Specifically, one year of data from the Terra MODIS surface reflectance product (MOD09GQ version 6), covering 1 July 2016–1 July 2017, was used. These MODIS images have a spatial resolution of 250 meters. NDVI layers were derived from the MODIS images and masked using the “state_1km” quality band to remove low-quality and cloud-contaminated pixels. For the PDF composite, ten layers were created, each representing a different NDVI percentile from 5%–95%, in 10% increments. For the sequential composite, a 12-layer product was generated by computing the median NDVI for each month. Further details on the classification methodology and datasets can be found in Khatami et al. 2020.

Two additional data products were used as ancillary data. The first dataset was the digital elevation model from the Shuttle Radar Topography Mission (SRTM; Reuter et al. 2007). The SRTM elevation data have one arc-second (about 30 m) pixel size. In addition to elevation, slope and aspect variables were calculated, from SRTM data, and used as three explanatory variables. The second ancillary dataset was the nighttime light data from Visible Infrared Imaging Radiometer Suite (VIIRS). Day/night band (DNB) from VIIRS are monthly-averaged nighttime radiance composites. The DNB data are filtered for stray light, lightning, lunar illumination, and cloud cover, before averaging. In this work, DNB data for seven months, from October 2016–April 2017, were used. The cf_cvg band, which documents the total number of observations used to create each pixel value, was used to filter out pixels with a low number of observations, which were mainly due to cloud cover. Specifically, for each monthly composite, pixels with less than four cloud-free observations were filtered. Then, a single layer was created based on the minimum per-pixel value from the seven-month data.

The Landsat 8 data were used as the main remote sensing source, and the classification was performed at 30 m pixel size. Consequently, all the explanatory variables were resampled and re-projected to 30 m and the WGS84 coordinate system. In total, 102 explanatory variables were used for the 2017 classification.

Explanatory variables of the 2006 classification

Similar explanatory variables to the 2017 classification were used for the 2006 classification. Landsat 7 Collection 1 Surface Reflectance products were used for time series composting. Because of the Landsat 7 scan line corrector (SLC) error, Landsat 7 images have large numbers of missing pixels. Therefore, instead of a two-year time series, which was used for the 2017 classification, a four-year time series was used for the 2006 classification. Specifically, Landsat images from 1 January 2004–31 December 2007 were used. Five NDIs were calculated using the six spectral bands of Landsat 7 and were masked based on the f-mask cloud score. The PDF, sequential, and sinusoidal composites were created similar to the 2017 classification. This resulted in a PDF composite with 25 layers, a sequential composite with 20 layers, and a sinusoidal composite with four layers. The textural layers were not calculated using the Landsat 7 images because of frequent missing pixels due to SLC error. For the MODIS time series composites, MOD09GQ data from 1 January 2006–1 January 2007 were used. The MODIS NDVI composites were created similar to the 2017 classification with 10 and 12 layers for the PDF and the sequential composites, respectively. Finally, the elevation and its derivatives from SRTM were used as three ancillary explanatory variables. The VIIRS nighttime light data do not exist for 2006 and, consequently, were not included in the 2006 classification. In total, 74 explanatory variables were used for the 2006 classification.

Reference data of the 2017 classification were collected at 30-meter pixel size. The reference pixels were sampled based on stratified random sampling. Two existing land-cover maps of Ethiopia, the Copernicus Global Land Operations (https://land.copernicus.eu/en/products/global-dynamic-land-cover) and the GlobeLand30 (Chen et al. 2015), were used to determine strata (see Khatami et al. 2020 for more details). Overall, 1810 reference pixels were sampled for the 2017 classification. The reference label of sample pixels was manually determined through visual inspection of high-resolution images from 2017 on Google Earth. The same reference pixels (i.e., locations) were used for the 2006 classification. The reference label of pixels was collected based on existing historical high-resolution images on Google Earth. Because high-resolution images were not available for many locations during 2006, high-resolution images from 2004–2007 were used for reference data collection. However, high-resolution images were not available for many reference pixels even for the four-year period. Therefore, additional reference pixels were collected randomly from sparse areas, depending on the availability of high-resolution images over those areas. Overall, 1358 reference pixels were collected for the 2006 classification.

A machine learning-based random forest classification algorithm was used to perform the classifications. For each date, the reference pixels were divided randomly to training and test datasets. About 70% of the reference data were used for classifier training and the remaining 30% pixels were used for classification accuracy assessment and area estimation. The random forest classifier was tuned through a grid search to optimize the number of variables per split and the minimum number of samples at a leaf node for each of the classifications based on classification out-of-bag error (30% of training data). For the number of variables per split, values between the square root and two-thirds of the total number of variables were tested. For the minimum number of samples at a leaf node, values between one and five were tested. The number of trees for both classifications was set to 100; although a greater number was tested, accuracy was not improved. The final classification of each date was created using the optimal values of the two parameters.

Accuracy assessment and area estimation

The common practice to evaluate the accuracy of land cover classifications is through estimation of a population error matrix (population refers to the entire collection of map pixels) and accuracy measures. For a pixel count population error matrix with rows and columns corresponding to the mapped and the reference classifications, the cell value N_ij represents the number of pixels, for the entire map, classified as class i with reference class j (i & j ϵ {1, ..., K}, where K is the total number of classes). Marginal totals N_k+ and N_k+ represent the total number of pixels classified as class k and the total number of pixels with reference class k, respectively. The pixel count population values can be converted into proportions by dividing by the total number of map pixels (N). For example, p_ij = N_ij/N denotes the proportion of the map area that is classified as class i with reference class j. Similarly, p_i+ = N_i+/N and p_+j = N_+j /N denote the map and reference proportion of class i, respectively. The cell and marginal values can be used to calculate accuracy summary measures such as overall and class-specific accuracy. Note that knowing the true population error matrix values requires the reference label for the entire map pixels. Thus, in practice, the population error matrix is estimated using a sample of reference data.

Accuracy assessment of the produced land cover maps was conducted using the test dataset of each period, i.e., 30% of reference data of each period that were not used for classifier training. The reference data for each period were collected based on a stratified random sample, with strata defined by the overlapping of the Copernicus Global Land Operations and the GlobeLand30 products. Because the stratification sampling’s strata were different from map classes, the method defined by Stehman (2014) was used for error matrix estimation. To estimate a given cell proportion of error matrix (p_ij ), an indicator function was defined as follows:

(3)

The cell proportion p_ij is the mean value of y_u for all map pixels Ῡ. Therefore, an unbiased estimate of Ῡ can be obtained using the stratified sampling estimator, as follows:

(4)

Here, Ȳ̑ is an estimate of p_ij, H is the number of sampling strata, N_h is strata size (number of pixels in stratum h), and ȳ_h is the sample mean of y_u for stratum h. Eqs. 3 and 4 were used to estimate the cell proportions for all cells of the error matrix of each classification. The estimated error matrices were used to calculate the classification overall accuracy and class-specific user’s and producer’s accuracy.

The produced land cover maps were used to estimate the land cover area and the gross land change over the study period. The area of land cover classes or changes can be calculated directly from the classifications based on pixel counting. However, this could lead to substantial area estimation errors because any classification could include misclassifications (Czaplewski 1992, Gallego 2004, Stehman 2005). Specifically, map pixel counts correspond to the error matrix row marginal totals (N_k+), i.e., the number of pixels with map classification as class k, whereas the objective of area estimation is to estimate the true (or reference) area that is the column marginal totals (N_+k), i.e., the number of pixels with reference classification as class k. The reference land cover area (or proportion of area) might be estimated using a sample of reference data based on a probability sampling method. This means the same test dataset used for the accuracy assessment of each classification can be used for area estimation. The column marginal totals from the estimate error matrices (͡p_k+) provide unbiased estimates of reference class proportions (p_k+). However, the column marginal totals are estimated solely based on the sample data reference labels. This means the classified map would not be incorporated in area estimation and the area estimations could have large standard errors. Therefore, the model-assisted difference estimator (Stehman 2013) was used for area estimation. Because sample data were collected using stratified random sampling, land cover areas were estimated per stratum using the difference estimator and aggregated to obtain the total areas (see Stehman 2013 for estimator formulas).

Land cover trends at documented LSLA

The land cover products were used to evaluate two known LSLA datasets that are available for Ethiopia. This includes the LM database and the Ethiopian Land Bank (Fig. 1). The LM database is a collaborative effort to systematically document and verify information related to global LSLAs (Anseeuw et al. 2012). The database draws on a variety of sources to identify LSLA locations and includes useful ancillary information, such as intended use, deal size, and operational status. For this study, we filtered the data based on the target country (Ethiopia) and removed points with duplicate geographic coordinates or inadequate spatial accuracy. Data acquired from the Ethiopian government’s Land Bank consisted of 735 transactions in the western lowlands. Whereas the government records include the boundary of transactions, allowing for delineation represented as polygons, the LM records indicate the centroid point of transactions rather than their boundaries. Therefore, a buffer zone was constructed for each LM transaction point based on the deal size and was then used for calculating changes in cultivated area to compare information from the LSLA datasets.

To assess the LCC potentially associated with transactions, the Land Bank data were used to examine trends at LSLA parcels and their surrounding land. The 735 parcels were dissolved to create single features from transaction sites that share borders, resulting in 372 parcels. Additionally, 1 km buffers were created for the 372 transactions, and a clip process was used to reduce inaccuracy caused by overlap of buffers and other LSLA sites. The proportion of stable land and LCC was quantified using the land cover products, results from which were used to compare land cover trends among three regions of interest (ROI): Ethiopia, government documented LSLA, and the 1 km buffer surrounding LSLA. This enabled a better understanding of the effects created by LSLA and the potentially resulting shift in land cover.

As previously mentioned, modifications in land use often lag the change in land tenure following a transaction, though both changes may result in changes to land cover. To better examine LCC at sites that have exhibited modified land use following a transaction, a subsample was created from the 372 government documented transactions sites that only included plots where land cover changed to cultivated. This was based on LCC to cultivation between 2006 and 2017 that was greater than or equal to 5% of the area of a given LSLA. This generated subsamples of 87 LSLA sites where conversion to cropland was initiated, excluding two parcels that contained no data for 2017 because of cloud cover. Trends in LCC were then measured for the subsample of converted LSLAs and their buffers to assess the types of LCC in and around LSLA that have begun implementation.

RESULTS

Land cover classification

The produced land cover maps for 2006 and 2017 are presented in Fig. 2 (a–b). Stable and changed cover were identified based on the per pixel comparison of the classifications (Fig. 2, c–d). Savanna cover was the most common class for both years covering eastern, southern, and northwestern parts of Ethiopia. Cultivated land mostly occurs in the central and northern parts of the country, and an increase in the cultivated area was observed from the land cover maps (Fig. 2 a–b). The maps show some areas of deforestation in central regions of the country, and stable bare soil cover appears in the northeast. Most of the changed areas are located in the central, southwestern, and northern parts of the country (Fig. 2 c–d). However, some of the changes might be attributed to classification error, and it is necessary to adjust area estimations accordingly.

Accuracy assessment results for the two land cover maps are presented in Table 2. The overall accuracy for the 2006 and 2017 classifications was 77.36% and 76.60%, respectively. The savanna class had relatively higher user’s accuracy (UA) and producer’s accuracy (PA) for both years compared to the other classes. The cultivated area was mapped more accurately in the 2017 classification compared to the 2006 classification. UA and PA of cultivated classes were larger by 7.91% and 21.66% in the 2017 classification compared to the 2006 classification, but the dense forest class was mapped more accurately in the 2006 classification. The omission error for the dense forest class in 2017 was relatively large. This again emphasizes the importance of accounting for classification error when estimating land cover and land cover change area.

Fig. 3 shows the types of change indicated by the per pixel comparison of the two land cover maps. The changes for the urban/built and the water/wetland classes were small compared to the changes for the other classes (see Table 3) and so were excluded from Fig. 3 to enhance visualizations. The diagonal maps in Fig. 3 show the stable cover (between the two maps) for each class. The largest change was observed for the conversion of savanna to cultivated class. Several areas from the northwest to the southwest of the country changed from savanna to cultivated cover. Also, some conversions from dense forest to cultivated were detected in the southwestern areas of the country. Substantial changes were observed between dense forest and savanna classes. Pixels revealing a conversion from savanna to dense forest cover, potentially representing tree crops planted as a result of LSLA or intensification, were dispersed over central Ethiopia and toward the west. In contrast, conversion from dense forest to savanna was concentrated in the west and was possibly a consequence of forest degradation through, for example, charcoal production or deforestation and then a resultant conversion to a savanna cover type (i.e., less than 70% tree cover). Some areas in the northeast were identified as changed from savanna to bare soil, and some areas in the southeast were highlighted as changed from bare soil to savanna.

Land cover areas from the classifications were adjusted for classification errors using a model-assisted difference estimator previously explained. The adjusted area estimations are presented in Table 3. For the cultivated land, an increase in the area estimation from 149,353 km² in 2006 to 229,864 km² in 2017 was observed. The area of dense forest class also increased, from 105,524 km² to 196,367 km². Savanna cover decreased from 745,668 km² to 619,461 km², mostly because of conversion to cultivated and dense forest cover, illustrated in Fig. 3. Bare soil area also decreased, from 132,764 km² to 91,415 km². The most observed change for the bare soil class was the conversion to savanna cover.

LCC and LSLA

Comparison of existing datasets

The land cover products were related to the LM and Land Bank datasets to evaluate the accuracy of existing LSLA datasets. Calculations were made to quantify the proportion of cultivated area and the change in the proportion of cultivated area for each transaction record using the land cover products. Fig. 4a–b shows the histograms of the proportion of cultivated area for the LM records in 2006 and 2017, respectively. A large number of the transaction records had a small proportion of cultivated cover, with only a slight increase by 2017. As Fig. 4c shows, the change in the proportion of cultivated cover was small for many transaction sites, and some sites even had losses in cultivated cover. For the government documented LSLA, cultivated cover and change to cultivated cover were even less than the LM records. In 2006, most of the transaction records had close to zero cultivated cover (Fig. 4d). Cultivated cover for those areas was barely higher in 2017 (Fig. 4e–f). However, the transaction sites were expected to have substantial cultivated cover or change to cultivated cover. Such observations raise concern that the information being made available to researchers is not adequate for thoroughly understanding land transactions as socio-environmental phenomena that impact social and ecological functions. Furthermore, it highlights the fact that much of the land that has been leased to investors was in a state of non-intensive use prior to the expected shift to commercialized agriculture, which has implications for the potential social and environmental impacts of LSLA.

Comparison of LSLA, buffers, and conversion status

Further analysis was carried out to examine the impact LSLA may have on the landscape by using the LSLA data from the Ethiopian government’s Land Bank. The LCC was compared across three regions of interest (ROIs), which represent LSLA sites, a 1 km buffer surrounding LSLA sites, and the remaining area of Ethiopia. Calculations for Ethiopia excluded land occupied by LSLA and the buffers; however, this did not change the general national level LCC trends described above. Results of LCC in these three ROIs are presented in Table 4, which shows the stable and changed land cover as a percentage of the total area for each ROI. Stable land cover represented the highest proportion of land cover trends across all regions. More than 75% of the area analyzed persisted as savanna or forest between the two classification dates. The largest LCC for all ROIs came from conversion of savanna to forest or cultivation. Overall, there was a lower percentage of conversion to cultivated land cover at LSLA (4.2%) and the surrounding land (5.7%) compared to that exhibited across all of Ethiopia (12.4%). However, the LSLA parcels and their buffers returned higher proportions of LCC in the form of forest and savanna cover.

Trends in LCC at LSLA and their buffers were also evaluated for land acquisition sites where implementation was recorded (so as to reduce potential bias due to lack of implementation at a site). The last two columns of Table 4 highlight the area of each stable and change class found in land acquisition and buffer areas where this implementation has occurred. The values shown in Table 4 indicate the LCC as a percentage of the total area for LSLA sites and buffers, as implemented. Similar to the previous results, most land cover was stable, with the largest difference in LCC being savanna to cultivation in implemented land acquisitions (14.8%) and their buffers (20.7%). Another notable change among LSLA where implementation had occurred was the transition from forest to savanna, which was relatively high for both LSLAs and their buffers (4.8% and 5.8%, respectively) and for savanna to forest transitions (6.1% and 6.3%, respectively) for implemented land acquisitions and their buffers. Given all three of these rates are higher in the implemented land acquisition sites (Table 4, last two data columns) compared to the ROIs overall (Table 4, first three data columns), this does highlight how many areas of land acquisition have not yet been converted or the acquisition implemented on the ground, and so actual potential land cover change is likely to be much more significant than currently shown. This also seems to highlight the potential lag effects from transactions being documented versus implemented.

DISCUSSION

Ethiopia has undergone significant LCC from 2006 to 2017, with much of the change relating to increasing agricultural area and possibly related to LSLA. However, only a small portion of the exhibited change occurred within the boundaries of transactions, as illustrated by the initial review of the two available databases. The LM and Land Bank data did not consistently relate to areas of increasing cultivated land cover when matched with our national level, 30 m land cover classification. Although Keeley et al. (2014) noted the rapid rate at which land concessions have been granted in Ethiopia, the results of this study show a much slower rate of implementation and conversion, which complicates the ability of researchers to distinguish long-term effects of LSLA on the landscape. Despite the shortcomings of such a large-scale LCC product, comparison of the national level classification to high-resolution imagery highlights some limitations with the existing LSLA databases rather than the land cover product. These shortcomings have significant implications for the use of such databases to sufficiently understand and examine LSLA as a social-ecological process.

A visual inspection of the LSLA databases was performed using high-resolution historical imagery from 2006 and 2017 from Google Earth, which revealed the constraints of those datasets for identifying transaction activities. Such limitations can be mainly categorized as (1) areas identified as transaction sites that were not converted to cultivated land; (2) areas that appear as converted to intensive cultivated land on high-resolution imagery (and may therefore be LSLAs) missing from the LSLA databases; and (3) the boundary of LSLA records not matching with the actual conversion on the ground in many cases. Fig. 5 shows three such examples of LSLA records overlaid on high-resolution images from 2006 and 2017 and the change map produced in this research. Case A in Fig. 5 is an example of a transaction that appears to extend outside of the recorded LSLA boundary. However, the east side of the boundary for case A has not yet been converted to cultivation. Those changes were captured, to some extent, by the LCC map of this work. Case B in Fig. 5 is an example of a documented LSLA with boundaries that are in accordance with the actual change on the ground. Note that the areas changed to cultivated at the center of this example correspond to conversion to small-holder farms. Case C in Fig. 5 is an example of where land appears to have been converted to intensive cultivation in the high-resolution imagery, potentially for an LSLA, but is missing from the LSLA databases.

The LM database is an impressive collection of information concerning LSLAs at a global scale, particularly given the obstacles faced by those attempting to obtain governmentally controlled geospatial data regarding LSLAs. Although the LM in part relies on potentially uncertain regional and national datasets for their accuracy, it is clearly a tremendous source of information and provides much needed details regarding transactions. Yet limitations persist when considering the precise spatial location of the land clearings for LSLAs. Although the governmentally documented dataset lacks the useful ancillary information found in the LM database, it does maintain superior spatial attributes of LSLAs as compared to the LM. The strength of this research lies in the linking of the LM and government Land Bank databases to land cover changes at 30 m for the regional or national level, which can be applied to better assess LCC from land transactions and associated processes (Fig. 6). Land cover change analyses produced from remotely sensed data and utilized with large geospatial datasets to locate regions that have converted to large-scale agriculture allows researchers to develop a more spatially explicit and accurate recording of actual land cover change and its effects.

Using national-scale LCC maps, we were able to analyze all LSLA documented in the Ethiopian government’s Land Bank at a relatively high spatial resolution and across a large geographic extent. Results show that, compared both to Ethiopia overall and to the land surrounding LSLA transactions, LSLAs had a lower proportion of conversion to a new land cover type between 2006 and 2017 (Fig. 6). Much of the LCC across all ROIs consisted of transitions from savanna to cultivation. There was also a substantial amount of conversion from savanna to forest, though this may be caused by the implementation of agroforestry for smallholders or LSLAs, the latter of which are known to plant tree crops such as oil palm (Elaeis guineensis) and eucalyptus. However, without sufficient information from LSLA databases, it is difficult to accurately estimate the frequency of such circumstances. Additionally, whereas gains in forest cover was a prominent trend of LCC across all ROIs, there was a disproportionate change from forest to savanna in the buffers surrounding LSLAs.

To better understand the impacts that LSLAs and associated LCC have on the landscape, we examined results based on a subset from the government’s Land Bank dataset that indicated implemented land acquisition sites. Results revealed that approximately 20% of the LCC in implemented LSLAs was from savanna to cultivation (14.8%) or savanna to forest (6.1%). As would be expected, the LCC from savanna to cultivation was much larger for the implemented transaction sites. However, this trend remained consistent for the buffers as well. Regions near implemented transaction sites experienced LCC to agriculture at a much greater magnitude than land surrounding in the ROIs. This trend could illustrate several outcomes of LSLA implementation. It may be the consequence of smallholder displacement and the indirect land use change as these farmers establish new plots in the vicinity of their former farmland, which they can no longer access. Similarly, it may be indicative of outgrower schemes between local farmers and investors, whereby investors are now supplying sufficient inputs for smallholders to expand crop production. Regarding the conversion from savanna to forest, implemented LSLAs exhibited nearly twice the percentage of this LCC. The presence of an LSLA may drive this change in multiple ways. For example, the larger change to forest in converted LSLAs could be caused by more agroforestry but could also be due to the inability of local people to utilize resources from unconverted portions of these sites, thereby allowing vegetation to be reestablished. A similar pattern was exhibited for LCC of forest to savanna, where converted LSLAs and their buffers have a larger proportion of deforestation, which may be a precursor to more agricultural expansion. Interestingly, across all LCC types, it was the regions surrounding implemented LSLAs that showed the greatest change in land cover, suggesting that the initial ecological impacts of LSLAs may be exhibited on land external to the transaction site itself.

The uncertainty about the cause behind the LCC trends emphasizes the need for improved documentation and monitoring of LSLAs, particularly given their potential to drastically shape a new social and ecological order in their environment. Documentation of a national level LCC is critical for understanding and anticipating social-ecological changes brought on by LSLAs. As shown in this analysis, the majority of the land delineated by the LM and Land Bank sites resides in uncultivated areas (Fig. 6). Furthermore, consistent land cover mapping for the entire country would enable better assessment of potential outcomes from a change in land tenure and land use. It is well known that the location of LSLAs is vital in determining the magnitude of ecological impacts (Balehegn 2015). For example, with proper land management, a LSLA may benefit the landscape by increasing vegetation cover. Contrarily, clear-cutting forests to implement intensive agriculture will lead to environmental degradation. In addition to providing consistency for national matters of land management, our classification provides high-resolution data on land cover for such a large geographic scale of analysis. Previous studies that address LCC around LSLAs have been carried out at fine spatial resolutions and in some cases have distinguished smallholder and intensive farming. However, such studies are limited in geographic scope because of the rigorous work required for such detailed analyses (Shete et al. 2016, Williams et al. 2021), and large-scale analysis is needed to utilize LSLA databases to their fullest extent.

CONCLUSIONS

Land transactions are a social-ecological phenomenon that have gained much attention since the turn of the century, particularly in less developed countries. These transactions are encouraged by governments on the basis that they will provide improved wellbeing in host countries. However, this reality is yet to come to fruition for many people impacted by LSLAs. Although some case studies have shown LSLAs can enhance employment opportunities and can even result in improved ecological conditions, other instances have resulted in civil unrest and environmental harm (Balehegn 2015, Lay et al. 2021). Regardless of the social outcome, LSLAs represent a large-scale change on the landscape and can greatly alter human–environment interactions within a region. Efforts made by researchers to assess the consequences of LSLA implementation have been severely hindered by a lack of accurate geospatial data. For this reason, remote sensing remains invaluable for understanding various aspects of foreign investment in farmland.

This work highlights the need to more effectively document LSLAs and the resulting change to land cover. Specifically, it highlights the value in utilizing spatio-temporal geodata, which has been promoted as a valuable tool for assessing LSLAs and their impacts (Lazarus 2014, Lay et al. 2021). The results show disparities between areas of land cover conversion and the locations of LSLAs provided by global and national databases, with much of this land remaining uncultivated through 2017. Furthermore, it shows that the majority of the land occupied by LSLAs in western Ethiopia was initially covered by savanna and forest. This should raise concerns over the potential of frontier clearing of forests and savanna ecosystems, both of which are critical habitats for environmental sustainability.

Analyses were also carried out to assess LSLAs and the land surrounding investment sites, as represented by a 1 km buffer. Findings revealed that land adjacent to LSLAs experienced a greater proportion of LCC, particularly from savanna to forest or cultivation as well as from forest to savanna. This trend suggests that the initial ecological impact from LSLAs is exhibited on land that is external to the transaction sites, at least in regard to LCC. A similar trend was found when LSLAs were grouped by implementation status, with the surrounding areas displaying greater change than the LSLA parcels. As expected, the largest difference in LCC in implemented transactions was exhibited in changes from savanna to cultivated land. However, implemented transactions and their surroundings also experienced greater change from savanna to forest and forest to savanna. This trend can be explained by multiple potential drivers, such as smallholder displacement and clearing for new farmland. Similarly, regions with increased forest cover may be due to the production of tree crops like oil palm. Overall, the areas surrounding LSLAs currently appear to be more severely altered by LCC than the land now privatized by investors.

The national-level land cover and LCC maps produced in this study can be used to corroborate and expand the national and regional level studies that utilize the incomplete information currently available in global and national LSLA databases. A classification of this geographic scale offers advantages over more common local assessments of LSLAs because it enables a broader look at the effects of transactions. Furthermore, such maps may assist in spatially consistent monitoring of LSLA activity in Ethiopia at the regional and federal levels, a critical aspect of land management that has failed to be streamlined in the past (Keeley et al. 2014). Future work will require improvement in distinguishing cultivated land cover for smallholder use from the intensive agriculture indicative of LSLAs. Furthermore, advanced methodology is needed to generate denser time series of land cover characteristics, which can be used to identify the locations of LSLAs based on their distinct temporal signatures in spectral space. However, the product developed here, with a national level 30 m land cover classification across multiple dates, is already a momentous step forward in this regard, and of real importance to the management of LSLAs globally.

LITERATURE CITED

Anseeuw, W., L. Wily, L. Cotula, and M. Taylor. 2012. Land rights and the rush for land: findings of the global commercial pressures on land research project. International Land Coalition. Rome, Italy.https://grassrootsjusticenetwork.org/wp-content/uploads/2017/02/ILC-GSR-report_ENG.pdf#gf_227

Balehegn, M. 2015. Unintended Consequences: the ecological repercussions of land grabbing in Sub-Saharan Africa. Environment: Science and Policy for Sustainable Development 57(2):4-21. https://doi.org/10.1080/00139157.2015.1001687

Bustamante, M., C. Robledo-Abad, R. Harper, C. Mbow, N. H. Ravindranat, F. Sperling, H. Haberl, A. de Siqueira Pinto, and P. Smith. 2014. Co-benefits, trade-offs, barriers and policies for greenhouse gas mitigation in the agriculture, forestry and other land use (AFOLU) sector. Global Change Biology 20(10):3270-3290. https://doi.org/10.1111/gcb.12591

Chen, J., J. Chen, A. Liao, X. Cao, L. Chen, X. Chen, C. He, G. Han, S. Peng, M. Lu, W. Zhang, X. Tong, and J. Mills. 2015. Global land cover mapping at 30m resolution: a POK-based operational approach. ISPRS Journal of Photogrammetry and Remote Sensing 103:7-27. https://doi.org/10.1016/j.isprsjprs.2014.09.002

Czaplewski, R. L. 1992. Misclassification bias in areal estimates. Photogrammetric Engineering and Remote Sensing. 58(2):189-192. https://www.fs.usda.gov/rm/pubs_journals/1992/rmrs_1992_czaplewski_r001.pdf

Davis, K. F., M. C. Rulli, and P. D’Odorico. 2015. The global land rush and climate change. Earth’s Future 3(8):298-311. https://doi.org/10.1002/2014EF000281

De Maria, M. 2019. Understanding land in the context of large-scale land acquisitions: a brief history of land in economics. Land 8(1):15. https://doi.org/10.3390/land8010015

Debonne, N., J. van Vliet, A. Heinimann, and P. Verburg. 2018. Representing large-scale land acquisitions in land use change scenarios for the Lao PDR. Regional Environmental Change 18(6):1857-1869. https://doi.org/10.1007/s10113-018-1316-8

Deininger, K. 2011. Challenges posed by the new wave of farmland investment. Journal of Peasant Studies 38(2):217-247. https://doi.org/10.1080/03066150.2011.559007

Dixon, J., A. Gulliver, and D. Gibbon. 2001. Farming Systems and Poverty: Improving farmers’ livelihoods in a changing world. FAO and World Bank, Rome, Italy and Washington, D.C., USA. https://openknowledge.fao.org/server/api/core/bitstreams/49d65d89-37a4-4384-a9f6-eb470bc9e9a2/content

Federal Democratic Republic of Ethiopia, 2016. Growth and Transformation Plan II. 2014/15–2019/20. Volume I. Main Text. Federal Democratic Republic of Ethiopia, National Planning Commission, Addis Ababa, Ethiopia. Development. Addis Ababa.

Gallego, F. J. 2004. Remote sensing and land cover area estimation. International Journal of Remote Sensing 25(15):3019-3047. https://doi.org/10.1080/01431160310001619607

Gebru, T. D. 2016. Deforestation in Ethiopia: causes, impacts and remedy. International Journal of Engineering Development and Research 4(2):204-209. https://rjwave.org/ijedr/papers/IJEDR1602035.pdf

Interdonato, R., J. Bourgoin, Q. Grislain, M. Zignani, S. Gaito, and M. Giger. 2020. The parable of arable land: characterizing large scale land acquisitions through network analysis. PLoS ONE 15(10):e0240051. https://doi.org/10.1371/journal.pone.0240051

Keeley, J., W. M. Seide, A. Eid, and A. L. Kidewa. 2014. Large-scale land deals in Ethiopia: scale, trends, features and outcomes to date. International Institute for Environment and Development, London, England. https://www.iied.org/12575iied

Khatami, R., J. Southworth, C. Muir, T. Caughlin, A. N. Ayana, D. G. Brown, C. Liao, and A. Agrawal. 2020. Operational Large-Area Land-Cover Mapping: an Ethiopia case study. Remote Sensing 12(6):954. https://doi.org/10.3390/rs12060954

Lay, J., W. Anseeuw, S. Eckert, I. Flachsbarth, C. Kubitza, K. Nolte, and M. Giger. 2021. Taking stock of the global land rush: few development benefits, many human and environmental risks. Analytical report III. Centre for Development and Environment, University of Bern; Centre de coopération internationale en recherche agronomique pour le développement; German Institute of Global and Area Studies; University of Pretoria; Bern Open Publishing, Bern, Switzerland. https://www.doi.org/10.48350/156861

Lazarus, E. D. 2014. Land grabbing as a driver of environmental change. Area 46(1):74-82. https://doi.org/10.1111/area.12072

Liao, C., S. Jung, D. G. Brown, and A. Agrawal. 2020. Spatial patterns of large-scale land transactions and their potential socio-environmental outcomes in Cambodia, Ethiopia, Liberia, and Peru. Land Degradation & Development 31(10):1241-1251. https://doi.org/10.1002/ldr.3544

Messerli, P., M. Giger, M. B. Dwyer, T. Breu, and S. Eckert. 2014. The geography of large-scale land acquisitions: analysing socio-ecological patterns of target contexts in the global South. Applied Geography 53:449-459. https://doi.org/10.1016/j.apgeog.2014.07.005

Nalepa, R. A., A. G. Short Gianotti, and D. M. Bauer. 2017. Marginal land and the global land rush: A spatial exploration of contested lands and state-directed development in contemporary Ethiopia. Geoforum 82:237-251. https://doi.org/10.1016/j.geoforum.2016.10.008

Oya, C. 2013. Methodological reflections on ‘land grab’ databases and the ‘land grab’ literature ‘rush.’ Journal of Peasant Studies 40(3):503-520. https://doi.org/10.1080/03066150.2013.799465

Schoneveld, G. C. 2017. Host country governance and the African land rush: 7 reasons why large-scale farmland investments fail to contribute to sustainable development. Geoforum 83:119-132. https://doi.org/10.1016/j.geoforum.2016.12.007

Seleshi, Y., and U. Zanke. 2004. Recent changes in rainfall and rainy days in Ethiopia. International Journal of Climatology 24(8):973-983. https://doi.org/10.1002/joc.1052

Shete, M., M. Rutten, G. C. Schoneveld, and E. Zewude. 2016. Land-use changes by large-scale plantations and their effects on soil organic carbon, micronutrients and bulk density: empirical evidence from Ethiopia. Agriculture and Human Values 33(3):689-704. https://doi.org/10.1007/s10460-015-9664-1

Stehman, S. V. 2005. Comparing estimators of gross change derived from complete coverage mapping versus statistical sampling of remotely sensed data. Remote Sensing of Environment 96(3):466-474. https://doi.org/10.1016/j.rse.2005.04.002

Stehman, S. V. 2013. Estimating area from an accuracy assessment error matrix. Remote Sensing of Environment 132:202-211. https://doi.org/10.1016/j.rse.2013.01.016

Stehman, S. V. 2014. Estimating area and map accuracy for stratified random sampling when the strata are different from the map classes. International Journal of Remote Sensing 35(13):4923-4939. https://doi.org/10.1080/01431161.2014.930207

Thiombiano, L., M. F. Fonteh, M. Malo, P. T. Gicheru, and S. Mukunda. 2017. Guidelines for sustainable large-scale land deals in Africa. Food and Agriculture Organization of the United Nations, Tunis, Tunisia. https://farmlandgrab.org/post/view/27641.

Viste, E., and A. Sorteberg. 2013. Moisture transport into the Ethiopian highlands. International Journal of Climatology 33(1):249-263. https://doi.org/10.1002/joc.3409

Zhu, Z., and C. E. Woodcock. 2012. Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sensing of Environment 118:83-94. https://doi.org/10.1016/j.rse.2011.10.028