Rapid expansion of irrigated agriculture in the Senegal River Valley following the 2008 food price crisis

The expansion of irrigated croplands throughout the 20th century boosted global agricultural productivity, yet limited improvement occurred in sub-Saharan Africa where many irrigation schemes and policies under-delivered. We mapped the distribution of croplands under active irrigation between 1986 and 2020 for one of Africa’s largest and most important transboundary river basins—the Senegal River Valley; using Landsat imagery with a random forest classifier and hidden Markov Model. We document two distinct epochs of irrigation development. Initially, a period of stagnation where less than 900 ha yr−1 was added, lasting until 2008. Followed by a boom phase of rapidly expanding intensively irrigated production with ∼9,000 ha yr−1 added for the last 12 years. These epochs overlap with national agricultural policy frameworks: the 1980s laissez-faire policies limited state involvement in agriculture and promoted Asian imports; followed by a more interventionist period focused on promoting domestic production following the food price crisis of 2008.


Introduction
In many regions, irrigation is crucial to buffer agricultural production against rainfall variability and enable year-round cultivation (Rufin et al 2018). Between the 1950s and early 2000s, irrigated croplands across the globe almost tripled in area with accompanying increases in agricultural productivity (Ray et al 2012, Siebert et al 2015. However, the expansion and intensification of irrigated agriculture has not occurred uniformly around the world. In sub-Saharan Africa, only an estimated 6% of croplands (roughly 13 million hectares) are actively irrigated compared with 37% in Asia (You et al 2011). As a result, crop yield gaps in sub-Saharan Africa are some of the largest in the world; contributing to high levels of chronic poverty, food insecurity, and malnutrition across the region (Barrios et al 2008, Van Ittersum et al 2016.
Knowledge of the current and historic dynamics of irrigated croplands is critical to guide sustainable intensification and development of irrigation across sub-Saharan Africa, and other regions with underexploited water resources. Identifying areas where water resources are under-developed enables precise targeting of agricultural policy to improve water access and management practices, and, in turn, address challenges of chronic food insecurity, poverty, and climate vulnerability (Allen et al 2011, Jaafar andAhmad 2020). Similarly, spatio-temporal data on irrigated cropland dynamics can support the monitoring of existing irrigation projects, ensuring projects deliver on their stated goals and, where necessary, guide corrective or remedial action-e.g. rehabilitation of schemes, refinements to future project designs (Deines et al 2019(Deines et al , 2020. Such information is particularly important in sub-Saharan Africa, where past irrigation projects have delivered on average 20% of proposed production areas, despite higher investment costs than South Asia (Bjornlund et al 2020, Higginbottom et al 2021. Despite the importance of spatio-temporal monitoring currently available data remain limited, particularly in sub-Saharan Africa and other parts of the Global South (Venot et al 2021). National and global agricultural statistics typically quantify land use, including areas under irrigation, aggregated to administrative or country scales, with limited updates and no disaggregation between cropping seasons (e.g. rainy versus dry season). Irrigation statistics in sub-Saharan Africa are known to suffer from several inaccuracies and uncertainties that limit their use for tracking irrigated area dynamics and trends, most notably the omission of informal or unofficial irrigated areas that exist outside of official registries and reporting (Woodhouse et al 2017, Balasubramanya andStifel 2020). Similar issues exist with data collected by large-scale household surveys such as the World Bank's Living Standards Measurement Study (LSMS), which covers only a selection of countries, provides only infrequent repeat data collection, and captures very limited (if any) information on irrigation practices that cannot be linked to specific plot locations.
Satellite remote sensing provides a potentially powerful tool to overcome gaps in existing monitoring strategies; particularly in sub-Saharan Africa where technical and financial resources for in-situ monitoring are highly limited. While a range of satellite-derived irrigation maps exist (Biradar et al 2009, Thenkabail et al 2021, these are unsuitable for monitoring dynamics in sub-Saharan Africa. Data typically are too coarse to identify reliably smallholder farms that dominate agricultural landscapes, and commonly cover a single year or selected snapshots in time, limiting capacity to support analysis of temporal dynamics of water management policies and investments. More recently, studies in other regions have demonstrated continuous mapping of expansion and intensification of irrigation using machine learning analysis of stacks of moderate-resolution Landsat-derived annual spectral metric composites (Deines et al 2019, Rufin et al 2021. However, to date, these applications have focused on nations dominated by large-scale agriculture. In contrast, attempts to quantify spatial and temporal irrigated area dynamics in sub-Saharan Africa have been limited despite clear demands for such data from governments, donors, and policymakers. Here, we leverage freely available Landsat imagery and cloud computing resources to map active dry season irrigation at near-annual frequency between 1986 and 2020 for one sub-Saharan Africa's largest and most important river basins-the Senegal River Valley (SRV). Our approach uses novel time series analysis methods to quantify historic spatial and temporal rates of irrigation expansion and intensification in the SRV, combining Landsat metrics, random forest classification, and a hidden Markov Model to overcome challenges posed by gaps in Landsat archive that are widespread across SSA. This methodology is applicable to other localities where irrigation activity is dynamic but poorly quantified. Our analysis highlights two distinct epochs of irrigated agricultural development in SRV: stagnation with extremely limited expansion lasting until roughly 2008, followed by boom of rapidly expanding intensively irrigated production areas over the past decade. We discuss these findings in the context of large-scale water infrastructure developments, international conflict, and shifting political economy frameworks in the SRV, in particular demonstrating the important role of shifting national agricultural politics following the food price crisis of 2008. Our findings highlight the importance of state capacity and support to promote irrigation expansion.

Study area
Our study area focuses on the lower SRV (figure 1). The Senegal River is the 9th longest in Africa, flowing for 1086 km from the Fouta Djallon highlands of Guinea, through Mali, then demarcating the Senegal-Mauritania border and reaching the Atlantic Ocean. The SRV's headwaters are in the humid-Sahel, with around 1000 mm year −1 of rainfall; while the lower reaches are more arid, receiving 300-500 mm year −1 of rainfall in the July-October rainy season. In contrast, dry season rainfall (November-June) is minimal, with less than 3 rainy days per month. Dryseason agriculture depends on irrigation from the river and its tributaries, producing primarily rice with supplementary vegetables including tomatoes and onions. Agricultural policy in the SRV has reflected African political economy. The mid-20th century was characterised by optimistic belief in the power of infrastructure developments, especially large dams, to drive economic growth (Mold 2012, Awojobi andJenkins 2015). Across the Sahel, focus on dams and irrigation intensified with the droughts of the 1960s and 70s, when failing harvests drove major food insecurity and famine in the region (Woodhouse and Ndiaye 1990). In the SRV, a multilateral, Senegal River Basin Development Authority (in French: Organisation pour la mise en valeur du fleuve Sénégal, OMVS) proposed two co-operational dams to enable large-scale irrigation. In the river estuary, the Diama Dam would prevent saltwater intrusion; while upstream, the Manantali Dam and an accompanying 477 km 2 (11.3 billion m 3 ) reservoir would stabilise river flows for irrigated agriculture downstream, while providing hydropower. Both dams began construction in the early 1980s, but were not fully operational until 2001 as poor relations between Senegal Mauritania (including the 1989-1991 war) prevented co-operation. By the mid-1980s the World Bank and the IMF had withdrawn from large infrastructure, shifting to governance reforms through Structural Adjustment Programs (Moseley et al 2010). These reforms reduced state involvement, and severely curtailed agricultural outreach services. Following these reforms, West Africa became increasingly dependent upon food import from Asia, with domestic production stagnating (figure 2; Woodhouse and Ndiaye (1990)). More recently, food price spikes in 2008 (figure 2) and an increasingly unpredictable rainy season refocussed national and international attention on irrigation developments. In particular, Senegal initiated a series of policies to foster national selfsufficiency in rice by restarting subsidies for seeds, fertiliser, and irrigation equipment (Ministry of Agriculture and Food 2009) Further efforts focussed on improving management practises, ensuring timely harvests, and maintaining soil fertility and weed control (Tollens et al 2013).

Methods
Our analysis focuses on producing 27 annual dry season irrigation maps, covering the 1987-2021 period. Briefly, we generated these by constructing a nearannual series of Landsat composites. Next, training a random forest model for the year 2020, applying this to the full time series, and correcting the outputs using a hidden Markov Model.

Training data
We collected training data for a binary classification distinguishing between active irrigated cropland and an aggregate non-cropland class covering all non-irrigated land (fallow fields, grassland, urban etc). We used Landsat composites and Google Earth imagery to identify cropland under active irrigation during the 2020 dry season. In total, we collected 90 training polygons for cropland and 136 for noncrops, from which we sampled a random subset of 6000 pixels for each class for model training.
We also collected additional samples distinguishing areas that may be confused with cropland, such as wetlands and woodlands. These data were contrasted with a separate 'other' land cover class, which captured all additional vegetated and non-vegetated land cover types. In total, we digitised 156 wetland and 169 other polygons, and extracted 7500 random pixels from each class for model training.

Landsat processing and compositing
We accessed Landsat Collection 1 Tier 1 surface reflectance imagery via Google Earth Engine (Gorelick et al 2017). Images were preprocessed using LEDAPS (Landsat 4-7) or LaSRC (Landsat 8), and we masked clouds and cloud shadows using the F-mask layer Woodcock 2012, Dwyer et al 2018). Next, we calculated following vegetation indices: the Normalized Difference Veget- These indices are sensitive to vegetation growth stages and moisture content-indicative of active irrigation (Qi et al 1994, Xu 2006, Chandrasekar et al 2010. From the processed Landsat images we generated spectral-temporal metric composites. For the cropland classifications, we generated dry season composites of the median of each spectral band and vegetation index between January 1st and July 1st, for each year. The low number of observations prior to the launch of Landsat 8 in 2014 prevented the calculation of additional statistics. In total, we produced composites for 27 years between 1986 and 2021, with no or insufficient data available for the years 1989-1993, 1996, 1997, 1999, and 2009. For the permanent non-cropland area mask, we generated metrics using only Landsat 8 images from 2016-2020. We calculated a range of percentiles (10,20,40,50,60,80,95) and the standard deviation. To fully exploit the available image density we calculated these metrics over the dry season (January 1st-July 1st) and the full calendar year, resulting in 160 layers.

Classification, accuracy assessment, change dynamics
To produce active irrigation maps, we used our training data (section 3.1) and the median value Landsat composite for 2020 (section 3.2) to fit a random forest model, initiated with 500 trees. We applied this model to the other 27 composites, generating a timeseries of irrigation probability maps. A permanent non-cropland mask was also created using the training data and Landsat 8-derived metrics with a binary random forest classification, non-cropland areas were removed from further analysis. All land cover maps contain errors, which are propagated by comparing individual maps. Accordingly, we applied a hidden Markov Model (HMM) to the class probability layers (Abercrombie and Friedl 2015). This procedure discards changes where a transition is unlikely-e.g. an irrigation event located midway in 10 non-cropland years-based on a transition probabilities. Finally, we converted the probabilities to binary class labels, and removed cropland parcels of fewer than 9 connected pixels (0.81 ha) We quantified the accuracy of our classified maps using 1080 validation samples (40 per annual map) for active irrigation, with an additional 150 samples for areas never classified as actively irrigated. Each sample was visually inspected against high-resolution Google Earth imagery and a Landsat composite, and assigned either irrigated or non-irrigated. We calculated Overall, User's (Commission Error), and Producer's (Omission Error) accuracies using standard equations (Congalton and Green 2019). To quantify dry season irrigation dynamics, we calculated: (i) the first year of irrigation, (ii) how many years irrigation was detected, (iii) the proportion of postconstruction years actively irrigated.

Accuracy assessment
Our annual classified cropland maps obtained high accuracies for all years, with all overall accuracies greater than 0.95 (figure 3). The highest accuracy was achieved for four years at 100% accuracy (1987,1988,2017,2019), with the lowest in 1986 (0.95 ± 0.036). The slightly lower accuracy for 1986 is likely due to it being the first classification, and therefore not benefiting from temporal correction applied by the HMM, as there are no preceding years. On a class level (figures 3(b) and (c)), cropland was reliably classified in all years with producer's accuracies ranging from 0.82 ± 0.06 in 1986 to 1 for five years (1987, 1988, 2016, 2017, and 2018).

Spatial and temporal patterns of irrigation expansion
Our analysis identified large-scale, rapid expansion of new dry season irrigation in the SRV (figures 4(a) and 6(a)). The total area of dry season irrigation increased from a low of 4300 ha in the late 1980s to almost 70 000 ha by 2020, a sixteen fold increase (figure 6(a)). Expansion was evident in both nations: Mauritania increased from almost no dry season irrigation (less than 20 ha at the start of the time series) to 21 500 ha, with Senegal expanding from around 4000 ha to 50 000 over the same period. The rate of expansion was not consistent throughout our study period ( figure 6(b)). Prior to 2010, less than 1000 ha per year on average was added, predominantly in Senegal. However, post-2012 the rate of expansion drastically increased, with an average annual addition of 5000 and 3100 ha in Senegal and Mauritania, respectively (figure 6(b)).
Spatial patterns in the expansion process are evident in figure 4. During the 1980s and 1990s almost all irrigated crops were concentrated around Richard Toll in Senegal: characterised by intensively cropped sugar plantations with large field sizes owned by the Senegalese Sugar Company (SSC, Compagnie Sucrière Sénégalaise, figure 5). These agricultural land areas are cropped nearly every dry season in our observed record, with potential irrigation frequencies of greater than 0.75 ( figure 4(c)). Post-2010, new irrigated areas were added, initially close to the Senegal River (figures 4 and 5) and then gradually expanding outwards to surrounding areas. The frequency of cropping (typically around 0.5-0.75) in these more recently initiated areas of irrigated dry season agriculture is more variable than within the SSC plantations, but are generally high after accounting for variations in the onset year of dry season irrigation (figures 4(b) and (c)).

Discussion
We document the rapid, large-scale expansion of irrigated cropland in the SRV, with a 16-fold (65 000 ha) increase over the 1987-2020 period ( figure 6). This growth primarily occurred between 2010 and 2020, coinciding with the Senegalese Great Offensive for Food and Abundance (GOANA) agricultural expansion policy (Ministry of Agriculture and Food 2009). Throughout the 20th century, many African irrigation projects deteriorated rapidly or collapsed completely post-construction, with poor design and insufficient state capacity the most likely drivers (Adams 1991, Higginbottom et al 2021. In this context, the transformation of the SRV from a single harvest system with limited dry season cropping into an expansive double cropped zone is an impressive demonstration of state and infrastructure policy, contrasting with more pessimistic portrays of irrigation development in SSA more widely (Merrey and Sally 2017, Bjornlund et al 2020). Below, we contextualise our findings within the enabling political framework, detail possible lessons for irrigation planning, and discuss the role of Earth observation in mapping irrigation-led expansion and intensification of agriculture.

Stagnation and stability, 1987-2008
Between 1987 and 2007 the total area of dry season irrigation remained broadly static; only 10 000 ha was developed, less than 890 ha per year and almost exclusivity in Senegal (figures 4 and 6). Agricultural policy in this period was dominated by two key strands: an infrastructure policy L'Après-Barrage (after the dam) focussed on the construction of Manantali and Diama dams; and, the economic liberalisation agenda of Le Désengagement (disengagement) committed to rolling back state activities (Woodhouse and Ndiaye 1990). Both strategies promoted private investors and producer organisations involvement in irrigation management, with a more limited role for the state. However, implementation contributed to a multi-decadal Le Désengagement policies arose in the early 1980s, as part of overarching IMF Structural Adjustment Programme reductions in government spending (Schindler and Kanai 2021). In Senegal, the 1984 New Agricultural Policy reduced or eliminated subsidies for inputs (seeds and fertilisers), transferred to producer groups or privatised the assets and functions from the agricultural parastatal organisation, and established domestic price controls for cereals products (Woodhouse and Ndiaye 1990). Furthermore, a 50% devaluation of the CFA franc in 1994, attempted to stimulate domestic consumption. The aim of these policies was to incentivise private investors and farmers' associations to independently increase production, with domestic consumption motivated by a devalued currency (Moseley et al 2010). However, during the 1980s, input costs increased, maintenance declined, and little expansion or intensification occurred-with total crop production remaining broadly static (figure 2).
The Manantali and Diama dams were completed in 1988 and 1986, respectively, with Manantali generating hydropower by 2001. We show that completion of these dams did not increase dry season irrigation, contrary to promises during design and development. When planning began in the early 1970s, many similar projects were under construction across Africa, with donors and governments committed to large-scale irrigation schemes. However, by 1988 the era of large dams in Africa was over, the earlier boom produced many disappointing projects and donors soured on large infrastructure (Awojobi and Jenkins 2015). In particular, irrigation schemes attached to dams consistently failed to materialise as planned, with optimistic proposals and underwhelming returns (Higginbottom et al 2021). Furthermore, large-scale infrastructure investments were financed through low interest rates and readily available credit, which disappeared by the 1980's (Mold 2012).

Rapid growth, 2009-2020
Agricultural policies initiated in the 1980s across West Africa failed to increase domestic crop production, with imports from Asian markets increasingly required to meet demand (Moseley et al 2010). These imports became progressively more expensive throughout the 2000s and doubled in 2007-08, triggering social unrest and contributing to conflict in the region and elsewhere (Bellemare 2015). The price shock of 2007-08 refocussed attention on domestic production, with national food selfsufficiently returning to the political agenda of many African nations. In Senegal, the government aimed for food self sufficiency by 2017, announcing an ensemble of policies under the Great Offensive for Food and Abundance (GOANA) umbrella. Our analysis highlights the rapid returns from these policies, with more cropland added in 2013 alone (9700 ha) than in the previous 6 years combined (8900 ha). From 2013 onwards irrigated cropland coverage in both Senegal and Mauritania boomed, with a mean addition of 9500 ha per year.
The expansion of dry season irrigation in both Senegal and Mauritania over the last decade is a significant acceleration, compared to the latter years of the twentieth century. However, judged against their stated aims, policies and programmes were relatively unsuccessful. Senegal aimed for rice self-sufficiency by 2017, yet achieved a domestic production ratio of roughly 35% by 2020 ( figure 2). This target was to be achieved by developing 130 720 ha of irrigated rice with a cropping ration of 1.5, 115 720 ha of which would be located in the SRV. We identified a maximum dry season irrigated acreage of roughly 50 000 ha in 2019, which includes sugar and vegetable crops (figure 6). However, this discrepancy is more likely attributable to over optimistic planning than poor project delivery. Almost all large infrastructure projects over promise, under deliver, and over run in both time and budgets (Flyvbjerg 2007). Irrigation schemes are no exception, and compared to other sub-Saharan African irrigation developments the SRV has performed exceptionally well (Higginbottom et al 2021).
Several explanations exist for the relative successes of irrigation development programmes in SRV in contrast with other sub-Saharan African experiences. Firstly, large-scale, dry season irrigation in the SRV depends upon the Manantali and Diama dams; which were constructed and operational when the post-2008 expansion was planned. Irrigation developments, therefore, could be initiated immediately without being affected by delays in dam construction, which often result in postponements or reduce the available finance for the accompanying infrastructure needed for irrigation (e.g. canals). Accordingly, the main infrastructure deployed to facilitate irrigation expansion in the last decade were small diesel pumps and canals that sought to build upon and expand utilisation of already existing water storage and conveyance infrastructure (Ministry of Agriculture and Food 2009, El Ouaamari et al 2019). These component have two key benefits: firstly, they are modular, operating without pre-requisite facilities; secondly, they can be installed at multiple sites simultaneously, increasing the rate of deployment.
Secondly, the governments of Senegal and Mauritania adopted a whole-sector approach to agricultural development, not limiting action to infrastructure provision or water access (Ministry of Agriculture and Food 2009). Learning from mistakes of previous decades, the governments initiated input subsidies to farmers to reduce fertiliser and fuel costs (Ministry of Agriculture and Food 2009, Hathie 2019). For consumers, food prices were mediated by cash benefits, allowing farmers to obtain high sales prices to incentivise production. Finally, one aspect of the Le Désengagement policies that was retained and proved beneficial was a liberal approach to management. Many irrigated perimeters in the SRV have been initiated and operated by farmer village cooperatives, supported by state agencies (El Ouaamari et al 2019, Brosseau et al 2021. The benefits of inclusive management, over authoritarian styles, have been observed on other large irrigation schemes, notably the Office Du Niger (Mali) and Gezira (Sudan) (Aw andDiemer 2005, Bertoncin et al 2019).

Role of Earth-observation
Our analysis demonstrates the potential for Earthobservation analysis to document spatio-temporal patterns of irrigation expansion in regions such as the SRV, where in-situ monitoring is currently sparse and likely impractical to expand further in the future. EO time series of irrigated croplands can shed light on the performance and success of irrigation development programmes and policies, helping guide and inform ongoing efforts to deliver improvements in agricultural water security and climate change adaption (Deines et al 2019, Rufin et al 2021. In the SRV, we provide compelling evidence for the rapid recent expansion of irrigated croplands driven by responses to global commodity price spikes, whilst also highlighting the overall failure of many irrigation projects in the region to deliver on originally stated proposals to expand dry season production areas. A key challenges for accurately mapping longterm agricultural change at moderate spatial resolution in Africa is the sparseness of the early Landsat archive, particularly before the 2014 launch of Landsat 8. Our approach, combining Landsat metrics, random forest classification, and a hidden Markov Model, demonstrates a solution to this issue and produces highly accurate results (figure 3). Unexpectedly, annual classifications were comparably accurate in the data-sparse 1980s as during the post-Landsat 8 years. We attribute the high accuracies in the early years of our study period to the dominance of commercial sugar plantations with clear perimeters and large field sizes. These plantations are a relatively simple classification task compared to smallscale, staple crops, which only became widespread in the SRV from the 2010s, by which point the Landsat archive is more dense and able to handle more complex classification tasks. Furthermore, the hidden Markov Model proved effective at removing miss-classified pixels, which typically occur for only a single year, while retaining genuine land cover changes. hidden Markov Models may therefore assist other land change studies in regions with sparse Landsat imagery availability (Abercrombie and Friedl 2015).
While our analysis demonstrates the strong potential for EO-based monitoring of irrigated cropland expansion in similar regions of sub-Saharan Africa, there are several limitations and caveats to the generalizability of our findings that are important to address in future work. Firstly, we are unable to disaggregate irrigated cropland into different crop types, due to the restrictiveness of the early Landsat archive. Maps of the initiation and distribution of different crops would be extremely useful from hydrologic, agronomic, and political economy perspectives. The dominant crops in the basin (rice, sugar, vegetables) have distinct roles in rural economies, while also having differing water requirements that may affect wider, basin-scale water resource sustainability (Deines et al 2020). Consumptive agricultural water use, and the associated water productivity of irrigated cropland, can be quantified via thermal Earth observation. Combining land cover maps with these types of data, such as the FAO's Water Productivity through Open access Remotely sensed derived data (WaPOR, (Food and Agriculture Organization 2018)), would enable assessment not only of changes in irrigated areas over time but also of variability in water use and water productivity across new and existing irrigated croplands. Such analyses would provide valuable insights about how agricultural expansion is influencing overall consumptive water use within a region and help to identify areas of lower water use productivity that require interventions or remediation to ensure limited water resources are used effectively and efficiently , Safi et al 2022.
Secondly, the 30 m spatial resolution of Landsat precludes closer analysis of specific highly localised small-scale irrigated production areas and systems. Such analysis could provide useful information on the contrasting roles of farmer-led or state-led developments, and the incentives for irrigation development and expansion. Our Landsat derived maps are unable to discriminate between these modes, but should not exclude the smaller developments. High-resolution imagery would be required to distinguish between development types, which entails a compromise between spatial and spectral resolution, however, recent studies developing Sentinel 2-Planet fusion products have strong potential in this area (Johansen et al 2022).
These limitations prevent us from conducting a more detailed analysis of dam-induced transitions from traditional methods of agriculture to more labour-intensive, industrial farming systems. For example, prior to the 1988 construction of the Manantali dam, agriculture in the SRV was dominated by recession farming of millet and sorghum on the river floodplain (Saarnak 2003, Degeorges andReilly 2006). Following impounding of the river, field studies reported reduced viability of traditional recession farming, due to a combination of reduced floodplains, lower or no annual flooding, and unpredictable deluges (Rasmussen et al 1999, Adams 2000, Saarnak 2003; issues that are typical of dam developments across mid-20th century Africa (Adams 1992). In response to loss of recessional farming opportunities, many farmers in the SRV adopted diesel pumps for extracting water, a comparably more expensive and labour-intensive approach (Rasmussen et al 1999). Our analysis will detect increased of dry season pumping irrigation, but is not tailored to identify abandonment of non-dry season recessional floodplains that may have occurred as a result of the development of infrastructure systems to support intensification of irrigated dry season farming in the SRV.

Conclusion
Irrigation development in sub-Saharan Africa has generated much commentary and debate. Proponents argue that large-scale irrigation is an essential prerequisite for economic growth and food security across the region. Meanwhile, the costly failures of previous projects drives much scepticism and pessimistic assessment of current ambitions for large-scale irrigation developments projects. Within this debate, the SRV represents an interesting case study. Policies initiated in the 1980s to expand irrigated dry season crop production were an unequivocal failure. We show that acreages remained persistently low even after two major dams were completed, with a large proportion of the proposed 375 000 ha of irrigated cropland expansion never materialising. Conversely, we document and shed light on the rapid expansions of dry season irrigation post-2010 driven by a range of state and donor-led policies enacted to accelerate domestic crop production in response to global commodity price spikes. However, this expansion was, on it is own terms, still a marked failure against proposed targets. This 'failure' highlights the challenges of reliably anticipating and delivering agricultural expansion and intensification through large-scale irrigation infrastructure projects and programmes, which routinely miss proposed targets even if they successfully achieve broader aims, such as increasing and modernising productivity. Overall, our study highlights the benefits of Earth-observation for improving assessment of irrigated development projects in regions such as sub-Saharan Africa with sparse insitu monitoring infrastructure and resources. These approaches, allow governments and donors to generate objective data at scale on the performance of existing and new irrigation projects, helping to guide and refine the design and targeting of investments that will be essential to meet goals of improving food security, climate adaptation, and rural economic development.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors.