Between 1970 and 2008, global irrigated area increased from 170 million to 304 million hectares (Vogels M. F. A. et al., 2019). In sub-Saharan Africa however, as little as 4–6% of cultivated area is irrigated, given the lack of electric grid infrastructure and the high cost of diesel (Wiggins et al., 2021). Locating isolated irrigation identifies areas that can support higher quality energy provision services—e.g., a grid connection or minigrid installation—as these sites can sustain higher energy demands and the attendant electricity costs (Conlon et al., 2020). Facilitated through informed planning, irrigation expansion has a direct impact on poverty reduction: In Ethiopia, one study found that the average income of irrigating households was double that of non-irrigating households (Gebregziabher et al., 2009).

In data poor locations, satellite imagery provides a source of detailed synoptic observations of irrigated agriculture (Pervez et al., 2014). A previous irrigation mapping effort in Ethiopia used three 1.5 m resolution SPOT6 images to distinguish between large-scale and smallholder irrigation in the Ethiopian rift (Vogels M. F. A. et al., 2019). This approach was then adapted to intake timeseries of 10 m Sentinel-2 imagery to predict irrigation presence across the horn of Africa (Vogels M. et al., 2019). While both studies demonstrated high accuracies over collected observations, limited labels precluded a more rigorous performance assessment over the entire area of interest. Other studies have used multiscale imagery to detect irrigation, including one that fuses MODIS and Landsat imagery to identify irrigated extent, frequency, and timing in northwestern China (Chen et al., 2018). Here, unique advantages of satellite imagery products at different resolutions are exploited: 250 m MODIS imagery is valuable for characterizing vegetation over large areas (Huete et al., 1999), while decameter resolution imagery from Landsat or Sentinel-2 missions can better discern plot extent (Phiri et al., 2020).

Deep learning techniques have become widely used for land process classification, as they uncover intricate structures in large, complex datasets (Lecun et al., 2015); and provide a robust method of handling phenological variability (Zhong et al., 2019). However, despite increasing availability of remotely sensed imagery, computing resources, and advanced algorithms for information extraction, high-quality labels remain scarce and expensive to acquire. Methods of overcoming label scarcity generally fall into one of four categories: 1) Using pretrained networks; 2) unsupervised and self-supervised learning; 3) data augmentation; or 4) additional label collection (Li et al., 2018). Even as pretrained networks like ImageNet (Deng et al., 2009) are highly effective for true-color image classification, these networks’ weights do not translate to tasks that intake multispectral or hyperspectral imagery (Tao et al., 2022). Unsupervised learning techniques, including those that ensemble different clustering methods—e.g., Banerjee et al. (2015)—have been shown to effectively organize unlabeled imagery. Existing work has also demonstrated that training a Generative Adversarial Network (GAN)—itself a type of unsupervised learning—has allowed for improved change detection performance on multispectral imagery, e.g., Saha et al. (2019). For data augmentation, three techniques are often implemented: Image translation, rotation, and flipping (Yu et al., 2017; Stivaktakis et al., 2019); however, these techniques do not have obvious analogues for pixel-based classification.

In assessing the impact of training dataset size on land cover classification performance, Ramezan et al. (2021) finds that investigating multiple types of classifiers is recommended, as the performance of specific classifiers is highly dependent on the number of training samples. A number of other studies have introduced methods for obtaining training samples, including collection via hand-engineered rules (Abbasi et al., 2015); normalized difference in vegetation index (NDVI) thresholding (Bazzi et al., 2021); finding neighboring pixels that are highly similar to labeled pixels (Naik and Kumar, 2021); and visual inspection of high-resolution (Vogels M. et al., 2019) and decameter resolution (Wu and Chin, 2016) imagery. Lastly, while larger training datasets generally yield better model performance, condensing input samples via dimensionality reduction has been demonstrated to increase land cover classification accuracy (Stromann et al., 2020; Sivaraj et al., 2022).

Another lingering issue in land process mapping is determining the conditions under which a model can be utilized in locations beyond where it was trained. Site-specific methods may not be easily transferable to other places or climes (Ozdogan et al., 2010; Bazzi et al., 2020), and the performance of transferred models can often only be assessed after full implementation in a novel setting (de Lima and Marfurt, 2020). Therefore, processes that yield insights about model transferability before training and inference offer benefits to researchers seeking to understand the maximum spatial applicability of their approaches.

As current methods primarily focus on already well-understood areas of interest with existing datasets, new techniques and products need to be developed for parts of the world lacking labeled data. In the realm of irrigation detection, new methodologies and mapping products can help identify locations for further energy system planning and investment, as these areas contain latent energy demands that can make higher quality energy services cost-effective and increase incomes. To this end, the following paper presents a multiscale methodology that leverages 250 m MODIS imagery for regional phenological characterization and 10 m Sentinel-2 imagery for irrigation detection on smallholder plots. This approach is then applied to the 205,000 km² Ethiopian Highlands, whereby it introduces a novel method of label collection; an evaluation of different classifier architectures and training strategies that ensure model applicability within the area of interest; and an assessment of irrigated area in the Tigray and Amhara states of Ethiopia for 2020 and 2021.

Identification of dry season greening as potentially irrigated agriculture must take into account spatiotemporal variations in native vegetation phenological cycles. The complex topography of the Ethiopian Highlands and East African rift system, combined with the latitudinal movement of the InterTropical Convergence Zone (ITCZ) and seasonal upwelling of the Somali current in the Arabian Sea produces a diversity of rainfall patterns that control annual vegetation phenological cycles in the study area ¹ . In order to provide phenological context with which to identify anomalous dry season greening, a regional vegetation phenology map is derived from spatiotemporal analysis of timeseries of vegetation abundance maps. Using the spatiotemporal characterization and temporal mixture modeling approach given by (Small, 2012) applied to timeseries of MODIS enhanced vegetation index (EVI) maps, four temporal endmember (tEM) phenologies are identified that bound the temporal feature space of all vegetation phenology cycles observed on the East African Sahel. These four tEM phenologies form the basis of a linear temporal mixture model that can be inverted to provide tEM fraction estimates for each pixel’s vegetation phenology. Figure 1 presents a spatiotemporal phenological characterization for the country, created from 16-day 250 m MODIS EVI imagery between 1 June 2011 and 1 June 2021.

The four tEMs extracted for Ethiopia are as follows: a single cycle tEM, representing a single annual vegetation cycle per year that peaks in September/October; an evergreen tEM, representing perennial vegetation; a double cycle tEM, representing semiannual vegetation cycles observed on the Somali peninsula; and a non-vegetated tEM, representing barren or non-existent vegetation. The ensuing phenology map in Figure 1 contains unmixing root mean square (RMS) error less than 10% for 90% of the pixels; additional unmixing error statistics and the locations of the extracted tEMs in principal component (PC) feature space are shown in Supplementary Figures S1, S2.

Figure 1 roughly divides into 4 quadrants. In the northeast quadrant, Afar appears as dark green, indicating that none of the 4 tEMs contribute significantly to phenologies in this part of the country: The vegetation that does exist in this mostly barren area is represented by low levels of evergreen tEM abundances. In the southeast quadrant, dominated by Somali and a portion of Oromia, vegetation patterns cycle twice annually. This is an area with bimodal rainfall but low total annual precipitation that results in the double cycle tEM containing peak vegetation abundances lower than those of the single cycle and evergreen tEMs. It follows that southeast Ethiopia is more pastoral with sparser vegetation than other parts of the country.

The southwest quadrant—covering Southern Nations, Nationalities, and Peoples’ (SNNP) Region, Sidama, and the western portion of Oromia—contains significant amounts of evergreen vegetation, as is demonstrated by its bright green hue. Here, evergreen vegetation is supported by bimodal rainfall with higher levels of annual precipitation than in eastern Ethiopia. In contrast, the northwest quadrant of the phenology map contains red-dominant color gradients, indicating phenologies similar to the single cycle tEM. This portion of the country, known as the Ethiopian Highlands and comprising of Amhara and Tigray, is highly agricultural; the main cropping season lasts from June to October and coincides with the primary kiremt rains, with some secondary cropping following the lighter belg rains from March to May. Accordingly, cropping that occurs during the dry season between November and March is likely to be irrigated.

In presenting a map of dominant vegetation phenologies in Ethiopia, Figure 1 provides a guide for land cover classification applicability within the country. For instance, a dry season irrigation detector trained in Amhara will perform poorly in SNNP, as phenological patterns differ significantly across these states, and dry season crop cycles exhibit different vegetation signatures. In contrast, a dry season irrigation detector developed across Amhara can be transferred to Tigray or Benishangul-Gamuz, due to regional phenological similarities.

The named, italicized outlines in Figure 1 represent the 8 areas containing labels used in this paper, referred to as regions: The yellow outline indicates a region where labels were collected via a ground survey, and the purple outlines indicate regions where labels were collected by means of visual interpretation and timeseries inspection. Full information on the labeled data collection process is presented in Section 3.

The data collection portion of this paper’s methodology consists of pairing Sentinel-2 imagery with labeled polygons to train an irrigation detector. Here, a pixel timeseries paired with a binary irrigation/non-irrigation label constitutes a sample. Irrigation is defined as such: A pixel is irrigated if its phenology includes at least one non-perennial vegetation cycle during the dry season, 1 December to 1 April for the Ethiopian Highlands. Conversely, a pixel is non-irrigated if its phenology demonstrates only vegetation growth that can be attributed to the area’s known rainy seasons. Irrigated areas are only of interest if they contain dry season vegetation cycles; this strict definition of irrigation excludes supplemental irrigation practices and perennial crops that may be consistently irrigated throughout the year.

This paper makes a set of contributions to the literature for learning from limited labels. First, it demonstrates a process of collecting training data to supplement ground-collected labels that improves on previous methods of sample collection—such as using imagery from a single timestep or simple vegetation content heuristics—as it verifies the existence or non-existence of full vegetation cycles during the dry season. Second, an evaluation of inputs, classifier architectures, and training strategies is presented for achieving irrigation classifier applicability to a larger area. Results indicate that enhanced vegetation index (EVI) timeseries outperform a full set of spectral bands as inputs; that randomly shifting input timeseries prevents classifier models from overfitting to region-specific input features; and that a transformer-based neural network produces the highest prediction accuracies in unseen target regions. Due to the close similarity of performance metrics and alignment of predictions, the faster training, more easily interpretable CatBoost architecture is also shown as a suitable alternative for irrigation mapping efforts.

Prediction results indicate strong classifier performance over sample timeseries from regions not seen during training. On data from withheld target regions, transformer-based classifiers achieve mean F₁ scores above 0.95 when four or more regions’ data are included during training; using labels from all 7 visual collection (VC) regions, the transformer-based classifier achieves an F₁ score of 0.932 on the ground collection (GC) labels around Lake Tana. Over an independently collected set of more than 400,000 samples collected for performance assessment, the same classifier achieves 98.3% accuracy over non-irrigated samples and 95.5% accuracy over irrigated samples, demonstrating strong performance throughout the entire Ethiopian Highlands.

Deploying a transformer-based classifier trained on samples from all 8 label collection regions yields insight into changing irrigation patterns. Results suggest that from 2020 to 2021, irrigation in Tigray and Amhara decreased by 40%. In Tigray, this decline was most precipitous in the Northwest and Southern zones, which saw percent changes in irrigated area of −72.3% and −51.1%, respectively. The Western zone of Tigray was the only zone to see an increase in irrigated area from 2020 to 2021; even so, this increase amounted to 279 Ha in a zone with a total area of 1,331,652 Ha. Amhara is predicted to have had similar decreases in irrigated area: Apart from the Wag Hamra zone, which was predicted to have less than 0.08% of its area irrigated in 2020 or 2021, all zones in Amhara experienced a change in irrigated area between −25.0% and -82.3%. The largest declines by area occurred in North Shewa (−41,572 Ha), South Gondar (−29,636 Ha), and West Gojjam (−26,154 Ha). Combined, results for Tigray and Amhara predict severe reductions in dry season crop growth from 2020 to 2021, findings that align with recent reports of food insecurity following the eruption of civil conflict in Ethiopia in late 2020.

Despite presented performance metrics indicating high levels of prediction accuracy, there are a few limitations to the proposed methodology that are important to mention. First, the study area is limited to the Ethiopian Highlands, a highly agricultural, climatologically consistent area that is dominated by rainfed cropped phenologies. As the irrigation classifiers are only trained to separate dry season crop cycles from rainfed vegetation cycles—associating identified dry-season cropping with irrigation presence—they will perform poorly in settings with different rainfall and phenological patterns. Relatedly, the trained irrigation classifiers do not identify irrigation used to supplement rainy season precipitation, irrigation of perennial tree crops, evergreen vegetation in riparian areas, or irrigation that supports continuous cropping, as the phenological signatures of these types of vegetation are difficult to distinguish from evergreen, non-cropped signatures. This discrimination task is left for future work. Lastly, classifiers are trained only on cropped phenologies, which constitute a portion of the vegetation signatures that exist in the area of interest. To manage the other phenologies present at model inference, prediction admissibility criteria are implemented. Nevertheless, these criteria are imperfect: There are surely irrigated pixels which have been mistakenly assigned a non-irrigated class label, along with non-cropped pixels which have evaded the admissibility criteria.

While the presented methodology is applied only for the task of irrigation identification in the Ethiopian Highlands, the strategy of regional phenological characterization to provide context for geographically informed selection of training samples and model applicability can be transferred more broadly to a range of land process mapping objectives. The suitability of this approach in the field of machine learning with limited labels is supported by results comparing classifier architectures and hyperparameter choice to assess the question of result uniqueness that overshadows all land cover classifications. As discussed by Small (2021), what is presented as the map is often just a map—one of many different products that can be obtained from the same set of inputs with different classifiers and hyperparameter settings. By assessing multiple classifier architectures and quantifying prediction sensitivity, this approach demonstrates consistency in results and indicates the uncertainty that can be expected of the resulting irrigation maps; as such, it provides a process for building robust classifiers in settings with scarce labeled data.