Improved land monitoring to assess large-scale tree plantation expansion and trajectories in Northern Mozambique

Northern Mozambique (382,000 km²) is a relevant study area, with all of the major identified risks and benefits of tree plantation expansion. One of sixty-one signatories of the Bonn Challenge, Mozambique aims to increase its commercial tree plantation area to 1 million hectares by 2030 (Serzedelo de Almeida and Delgado 2019, The Bonn Challenge 2021). Current efforts focus on the northern provinces, which have been characterized both as a bastion of available land and a hotspot of large-scale land acquisitions (Di Matteo and Schoneveld 2016, Kronenburg García et al 2021). With the average farm size in Mozambique measuring less than two hectares (Smart and Hanlon 2014), and land titles for tree plantations averaging over 100,000 hectares, massive plantation expansion could threaten rural livelihoods and displace tens of thousands of smallholder family farms. However, tree plantation expansion can also benefit local communities, especially when communities successfully negotiate equitable compensation packages with private companies wishing to acquire usage rights of community lands. Such packages can include creating employment opportunities; building essential infrastructure; offering employee benefits such as housing, medical care, education; paying royalties; supporting small scale plantations; and establishing community development funds. When the (legally non-binding) compensation agreements are respected, tree plantation expansion can be a significant boost to local development (Bleyer et al 2016).

Most of the study area lies within Miombo ecoregions. Often subject to natural and anthropogenic disturbance, Miombo woody and herbaceous vegetation are characterized by their high natural recovery capacity after disturbance due to agriculture, charcoal production, logging, and fire (Jew et al 2016). Tree plantation expansion onto natural forest and grassland presents a risk of increasing net carbon emissions (Lewis et al 2019). Emissions from clearing natural forest might not be counteracted by an equal or larger amount of greenhouse gas emissions sequestered by tree plantations established on the same area of land. Given the high recovery rate of Miombo woody and herbaceous vegetation, it may be more efficient to increase tree cover through natural regeneration than by converting Miombo grasslands to tree plantations (Fleischman et al 2020). However, wood from tree plantations established on cropland and severely degraded woodlands could relieve pressure on more pristine areas as 86%–100% of households in the study area rely upon fuelwood – typically gathered from natural woodlands – for cooking (Woollen et al 2016).

Understanding the actual social, environmental and economic impacts of tree plantation expansion requires (among other things) comprehensive land monitoring that includes assessment of the type and location of new tree plantations and the use of the land prior to conversion to tree plantation. However, such information is lacking in Mozambique (Food and Agriculture Organization 2015). Past remote sensing assessments of land use have focused on miombo forests attributes and changes due to shifting cultivation, fire, and natural regeneration (Sá et al 2003, Sedano et al 2005, Marzoli 2007, Ribeiro et al 2008, Mitchard et al 2013, Ryan et al 2014). In the absence of precise information on tree plantation expansion, government estimates have been based on statistics derived from annual production plans and 'expert judgement' (Food and Agriculture Organization 2015). This is important in Mozambique as it aims to drastically increase its commercial tree plantation area from 75,000 hectares (the estimated area reported in Mozambique's 2015 Forest Resources Assessment report) to 1 million hectares by 2030. Current efforts focus on the northern provinces, our study area, where the government has identified millions of hectares of land suitable for tree plantations (Republic of Mozambique 2013). The land monitoring technique presented in this study help address this gap in knowledge of tree plantation extent and location.

Numerous remote sensing assessments have succeeded in mapping tree plantation extent and disaggregating it from natural forests (table 1). Yet, very few studies have also assessed the land use and cover before conversion to tree plantation (Echeverría et al 2006, Zamorano-Elgueta et al 2015, Heilmayr et al 2016, Poortinga et al 2019). This widely under-documented variable is critical to understanding the social, economic and environmental risks and potential benefits of plantation expansion. Remote sensing assessments of tree plantations in Africa area rare and none thus far have addressed land use prior to conversion to tree plantation.

Past remote sensing assessments of tree plantations can be divided into four groups of approaches (table 1). The first relies upon visual interpretation of very high resolution imagery to distinguish tree plantations from natural vegetation (Miettinen et al 2012, Petersen et al 2016, Curtis et al 2018, Kimambo et al 2020). This approach is effective but time consuming to implement and more susceptible to human error than other techniques. The second group uses traditional classifiers on multispectral imagery to interpret the spectral signatures of tree plantations relative to other features in the landscape (Li et al 2015). This approach has been successfully applied in Asia (Sun et al 2017, Xu et al 2018, Poortinga et al 2019), Central America (Fagan et al 2015) and South America (Zamorano-Elgueta et al 2015, Heilmayr et al 2016). The third group of studies utilizes intra-annual phenology-based algorithms, sometimes in conjunction with supervised classification algorithms, to identify tree plantations. Most of these studies focus on rubber plantations in East Asia where the plantation phenology differs greatly from that of agriculture and natural vegetation during the foliation and defoliation seasons (Dong et al 2012, Dong et al 2013, , Chen et al 2018). The final group also uses phenology, but places greater emphasis on inter-annual growth metrics and change events (Le Maire et al 2014).

Techniques that can detect plantations in one geography may not necessarily be effective in another. The main challenge to applying these approaches to Northern Mozambique is the landscape's diversity and intra-class spectral variability. Our study area spans six different ecoregions and includes a variety of natural evergreen and deciduous forest types that can appear similar to tree plantations (figures 1(E)–(J)). Furthermore, the spectral signature of tree plantations changes drastically with age, as newly established plantations resemble bare land, cropland or wooded grasslands, and mature plantation are easily confused with natural forests (figure 1, A-D) (Li and Fox 2012). Even across tree plantations of the same age, spectral signatures can vary widely with different species, soil types, nutrition, disease status, and levels of disturbance.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

A. Young eucalyptus, (−13.4468°, 35.6288°); B. Mature eucalyptus (−13.3270°, 35.3158); C. Young pine (−13.4360°, 35.6599°); D. Mature pine (−13.3431°, 35.3877°); E. Mango (−15.0475°, 39.14509°); F. Macadamia, (−15.4167°, 36.9095°); G. Eastern Miombo deciduous woodland (−13.4694°, 34.8642°); H. Southern-Zanzibar Inhambane Coastal semi-evergreen forest (−12.5013°, 40.3508°); I. Riparian evergreen forest (−11.691°, 37.3878°); J. Mangrove (−12.3931°, 40.5205°).

The locations mapped in the first panel of figure 1 are the same locations shown in ground-based photographs and satellite imagery in the second panel of figure 1. The rectangular ground-based photographs were acquired by the co-authors during field visits in 2017, while the circular satellite images are very high resolution satellite images from Google Earth centered around the (lat, long, WGS84) coordinates provided above with a radius of 100 meters. Google Earth attribute information and image acquisition dates are provided for each site.

Northern Mozambique is a large and varied landscape, approximately 382,000 km² in size. To assess land use changes throughout this vast area over a 17-year reference period, this study used all available USGS Landsat 7 and Landsat 8 imagery from the years 2000–2018, but focused its analysis on four points in time - i.e., 2001, 2006, 2012 and 2017 - that bookend the wave of foreign direct investments that began around 2008 and peaked in 2012. We accessed and processed the Landsat imagery in Google Earth Engine following steps, including cloud masking and median pixel composting, described in more detail by Bey et al 2020 and Flood 2013.

To train and test our classification algorithms, we contributed to and used a national dataset created for the SECOSUD project in collaboration with Sapienza University, Universidade Eduardo Mondlane (UEM), and Université catholique de Louvain (UCLouvain) (SECOSUD 2020). As several classes and time periods were under-represented in this dataset, we complemented it with our own land use and land cover (LULC) dataset produced by a joint UCLouvain-UEM team (Meyfroidt 2015). Both datasets used Collect Earth software, a free tool developed by the Food and Agriculture Organization of the United Nations to facilitate land assessment with freely accessible satellite imagery within Google Earth (Bey et al 2016). In Mozambique, this imagery included scenes acquired by Digital Globe (e.g. WorldView 3 and WorldView 4 satellites which have a 0.31 m spatial resolution, and QuickBird, which has a 0.65 m spatial resolution) and SPOT (e.g. SPOT 6 and 7 of 1.5 m resolution).

Certain classes in the study area, e.g. mangrove, other land, settlement and wetland have more consistent spectral signatures across the different ecoregions within the study area. Consequently, these classes were assessed with a single, nation-wide classification of each year of the reference period. In contrast, the cropland, forest, and grassland classes vary drastically within and between ecoregions. To optimize classification results for these classes, we conducted a series of supervised classifications using a spatially constrained subset of our training samples (Graesser and Ramankutty 2017). We divided the study area into water availability zones based on total annual precipitation and distance to wetland pixels derived from the national classifications. Eight classes were generated based on half quartiles of the total annual precipitation and the range of distance to water values. The supervised classification of each zone used training data from the target zone and the preceding and subsequent zones. The spatially overlapping results were resolved through a pixel voting technique in the which each pixel was assigned to the land use class with the highest number of 'votes' or occurrence within the pixel stack (Bruggeman et al 2016). Where an equal number of votes occurred for multiple land use class, we used a class hierarchy to determine the pixel land use class (Martínez and Mollicone 2012). In this manner, highland areas with high annual rainfall were assessed with floodplain areas where water availability for vegetation is similarly high. This is particularly important to minimizing misclassifications because Mozambique is characterized by a long dry season, during which foliation is largely controlled by water availability.

The tree plantation detection algorithm builds upon the work of Dong et al 2012 and Le Maire et al 2014 by exploiting the phenological and growth-based time series attributes that distinguish plantations from most other features within the Northern Mozambique landscape. We deviate from these two works by also using Object-based Image Analysis techniques that render our approach more accurate for our particular study site, and also more robust, flexible and adaptable to other regions of the world.

These previous works, on rubber plantations in East Asia (Dong et al 2012) and eucalyptus in Brazil (Le Maire et al 2014), have focused on the defoliation and foliation phase to distinguish deciduous rubber plantations from – usually evergreen or semi-evergreen - natural forests. In most parts of our study site, the situation is inversed, as miombo and mopane forests loose more foliage during the dry season than pine, eucalyptus and other tree plantations. We therefore applied a threshold to identify pixels with high dry season NDVI that may have a land cover of semi-evergreen or evergreen tree plantations. Tree plantations can potentially be distinguished from natural forests using only phenology in some areas; however, additional measures are needed when working over a vast and heterogeneous region. Other LULC classes in Northern Mozambique, such as coastal, riparian, mangrove, and montane forests are also evergreen and can have spectral and textural signatures identical to some tree plantations at any given point throughout the year. Thus, we used time series data over multiple years to better disaggregate tree plantations from natural, non-deciduous vegetation types. Prior to planting trees, commercial enterprises often clear the land of pre-existing vegetation, and for the first year or two after planting, the spectral and textural signature of tree plantations resemble that of cropland. In the subsequent years as exotic tree species grow and their canopy coverage approaches 100%, their NDVI and other vegetation index values grow steadily, unlike lands used for field crop, where index values rise and fall with each planting and harvesting cycle. The intra-annual NDVI values for nearly all vegetation types undulate through the wet and dry seasons each year, but for young tree plantations, subsequent wet, mid and dry season NDVI values are marginally higher than those of the preceding year. Therefore, we use the total NDVI – a sum of the mean value over all three seasons – as our second main feature for identifying tree plantations.

After delineating potential tree plantation objects using high NDVI and fast growth pixels, we applied a multiple criteria assessment to remove the remaining non-plantation patches based on object-level metrics. This step removed objects (1) smaller than 5 hectares, (2) with steep slopes, (3) with a low percentage of fast growth pixels, (4) containing a disproportionate amount of mangrove or (5) water pixels; and (6) with a high standard deviation. We also excluded objects in the province of Cabo Delgado. Although the province contains lands that are ecologically optimal for tree plantation, multiple sources indicate that there are no large-scale tree plantations in this province (Food and Agriculture Organization 2015, Di Matteo and Schoneveld 2016). The resulting tree plantation layers were superimposed on to the classifications for their respective years, replacing whichever class was determined by the classification algorithms.

This final step involving object-based image analysis is key to the flexibility of this plantation identification algorithm. Parameters targeting plantation surface area, gradual slopes, fast growth rates, and high dry season NDVI can be broadly applicable to other assessments of large-scale fast-growth tree plantations in Miombo or other semi-dry and sub-humid ecoregions. To apply this algorithm in other ecoregions or to target different types of plantations, the criteria can be adjusted to better reflect the local context.

Lastly, we made a spatial and quantitative comparison of our tree plantation data with corresponding information from government cadastral records. We obtained the spatial data associated with tree plantation land titles by submitting formal, written requests to the provincial governments of Nampula, Niassa and Zambezia. All through provided the records in Shapefile format. We processed the data in Google Earth Engine and compared it to our own remote sensing data generated in the same platform.

A description of the change detection, trajectory analysis, and accuracy assessment methods are provided in Supplementary Information 1.

An overview of the main land use dynamics in Northern Mozambique between 2001 and 2017 is presented in figures 3 and 4, while classification maps for the years 2001, 2006, 2012 and 2017 are provided in SI 4 along with a link to the accompanying interactive map. Over the 2001–2017 reference period, tree plantation area expanded by 175 ± 26 km². Our approach proved effective to map tree plantations and other land use classes with overall accuracies of ranging from 84% to 97.6% and the highest class accuracy for plantation was achieved with the 2017 classification, 95.9% (SI 5). Older maps achieved lower overall accuracies and lower tree plantation class accuracy due to higher commission error with forest, grassland and cropland.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Most new tree plantation land, 70%, was previously used for agriculture, while 30% was previously forest or grassland. The inset map of figure 4 provides a closer view of land use change in the Lichinga area, where the largest proportion of tree plantation expansion occurred.

We found major discrepancies between the spatial boundaries of tree plantation land titles and plantation expansion observed through remote sensing. Our results show that cadastral records can be inaccurate, imprecise, out of date, or simply nonexistent (figure 5). Only 59% of the 17,500 hectares of tree plantations identified in 2017 fall within land title boundaries, while 41% are on lands not currently designated for this land use.

Zoom In Zoom Out Reset image size

The total area covered by Northern Mozambique's tree plantation land titles is 831,000 hectares. The tree plantations identified within cadastral boundaries account for only 2.1% of the land allocated for plantations.