Abstract
Smallholder irrigation is crucial to livelihoods and growth in rural agro-economies. Identifying where it occurs provides valuable insights for agriculture extension, water management, infrastructure investments, marketing and energy provision. Remote sensing methods can aid but have been stymied by the small plot sizes, especially of irrigated plots. Humid tropical climates are characterized by high temperatures and heavy seasonal or year-round rainfall. This adds an additional layer of challenge due to fewer cloud-free imagery from persistent cloud cover. This study focuses on these settings to compare the performance, interpretability and transferability of classification using embeddings versus a feature engineering approach. Geospatial foundation models such as AlphaEarth Foundations (AEF) have emerged to address the contrasting scarcity of high-quality labels in comparison to increasing volumes of earth observation data. The embeddings from these models have been shown to generate maps from sparse labels outperforming traditional machine learning approaches in certain cases. In this study, we utilize a nation-wide smallholder irrigation survey data from Uganda as a data-rich region and pilot surveys from Zambia as a data-sparse region. We present a reproducible cloud-native Earth observation workflow in Google Earth Engine that discusses the tradeoffs and performance of the feature engineering approach compared to foundational model embeddings. The results show that a model trained on AEF embeddings outperforms a traditional feature engineering approach on a rich label dataset. The performance is similar in a data-sparse region. However, the embedding models do not lend themselves to auditable methods for applicability to regions where no data is available.
1 Introduction
Irrigation systems are vital to food security as they help meet crop water requirements and facilitate the production of additional cash crops during the dry season. A review by highlights the lack of significant growth in agricultural production in Africa compared to the rest of the world. While increases of more than two to four-fold have been observed across Asia, Europe, North America and Latin America since the 1960s, only a modest 70% increase is associated with Africa. Despite the stagnation in crop yields, Africa has seen a massive 250% increase in cultivated area since the 1960s compared to a 60% increase across Latin America or even a 20% decrease observed across North America and Europe. describes the agricultural policies that are focused on exports as the reason why food insecurity still exists in Africa. This focus has led to the neglect of investment in infrastructure, introduction of technologies and development of supply chains and markets.
Opportunities exist for expansion of smallholder irrigation in East Africa. However, access to reliable markets, irrigation technologies, credit services and poor water resource management are some of the key challenges to adoption of smallholder irrigation. Area under irrigation and thus agricultural productivity can be improved by adopting a water-energy-food (WEF) nexus perspective (). Smallholder irrigation schemes are often located away from any prevalent grid presence and rely on manual or diesel-powered pumps. The energy demand associated with pumping water and agro-processing present opportunities as a captive power. This combined with household demand can be used for planning a mini-grid and off-grid as the viable least cost solutions (). Irrigation loads provide a unique energy flexibility option by shifting the load during the daytime that complements the night time household loads associated with lighting, e-cooking etc. ().
African countries are characterized by infrequent census and surveys marring the ability to monitor the progress. A review by showed that the agricultural census in majority African countries is 10–20 years apart. The increasing availability of remote sensing data of varying temporal and spatial resolution paired with machine learning models can detect patterns and indicators of progress. () provides a review of different remote sensing methods utilized for mapping irrigated areas. This paired with increasing earth observation data provides an opportunity to reduce the overhead associated with national surveys by allowing for more targeted survey methods.
A number of large-scale geospatial foundation models have been released recently such as OlmoEarth developed by the Allen Institute for Artificial Intelligence (), TESSERA (Temporal Embeddings of Surface Spectra for Earth Representation and Analysis) from University of Cambridge (), Prithvi-EO-2.0 from IBM in collaboration with European Space Agency () and AlphaEarth Foundations (AEF) from Google (). These transformer-based vision models assimilate vast spatial, temporal and measurement contexts enabling production of classification maps from sparse labels. These foundation models have demonstrated state-of-the-art performance on a number of EO applications such as flood mapping, land cover mapping and crop mapping, etc. However, the use case for other problems that might not be as well-posed such as smallholder irrigation detection in Sub-Saharan Africa compared to traditional feature engineering methods remains unknown. This study aims at comparing the performance, interpretability and transfer learning capabilities of the two approaches.
In our previous work, we utilized a phenology map of the Sahel to characterize the Sub-Saharan African landscape into regions with single cropping cycles per year, dual cropping cycles per year and evergreen vegetation (). This is dependent on the onset and cessation of seasonal rainfall over Africa (). A dry season irrigation detection methodology developed for Ethiopian highlands was shown to be applicable in regions with single cropping cycles such as Northern Nigeria and Burkina Faso (). An extension of the work implemented the dry season irrigation methodology using Google Earth Engine tools to model temporal changes in irrigated areas (). However, these methods were only applicable to areas with a prolonged dry season. In humid environments such as Uganda, longer rainfall periods do not allow for vegetation to senesce thus rendering irrigated croplands indistinguishable from non-irrigated ones.
To address this limitation, this study considers two approaches. First is a feature engineering approach that incorporates topographic, climatic and anthropic variables in addition to the vegetation phenologies used in the earlier work. Second is to utilize embeddings from geospatial foundational models that compress vast amounts of earth observation datasets for semantic search and rapid classification tasks. In doing so, we compare and contrast the two approaches for smallholder irrigation detection in Sub-Saharan Africa. We define smallholder farms as plots that are of the order of a 10th of a hectare or less and often rely on surface water for irrigation (extracting water from swamps and dambos). A national scale survey dataset from Uganda serving as a data-rich region and pilot surveys from Zambia serving as a data-sparse region are used to quantify the performance of the two approaches. The main contribution of this work is a novel workflow to extend smallholder irrigation detection across humid tropical environments thus addressing the limitations of our earlier work () which was limited only to dry-season irrigation detection.
2 Materials and methods
This section provides a brief description of the survey dataset collected in Uganda and Zambia and the remote sensing features and datasets utilized in the study. Additional detail is provided in the Supplementary Material including a table of input features.
2.1 Study regions and survey data (Uganda; Zambia)
The first study region is Uganda which is characterized by biannual rainfall in the South in the crescent along Lake Victoria whereas the North receives annual rainfall with a short dry season (). A classification of irrigation schemes in Uganda by scale of irrigated area, water source and method of conveyance to the field is provided by . Smallholder farms are usually informal irrigation systems developed without planning and with little or no technical assistance. The majority of these systems are located on the fringes of swamps.
The second study region is Zambia where the smallholder farmers rely on dambos. Dambos are seasonally waterlogged areas found in the headwater zones of drainage systems or alongside streams. These areas provide grazing in the dry season. They also retain enough moisture to grow dry season crops using small-scale irrigation schemes (). Several studies provide an assessment of dambo hydrology in Southern Africa and estimate the density of dambos in Zambia to be 15%–20% of the total land area (; ).
A national scale survey designed to study irrigation among other productive uses of energy (PUE) was conducted in Uganda. Due to the focus on PUEs, a non-irrigated farm plot was only recorded within a sample area if no irrigated farm plot was found. This resulted in an imbalanced dataset of more irrigated (11,021) than non-irrigated (3818) farm plots. Similarly a pilot survey was conducted in Zambia to collect PUE data around proposed Rural Electrification Master Plan (REMP) sites (). The plot coordinates from the survey were then used with high resolution satellite imagery to draw polygons around irrigated farm plots resulting in 217 irrigated polygons across Zambia. Similarly, a total of 109 non-irrigated polygons were generated covering a total area roughly similar to the irrigated polygons. More details of the survey is provided in Section 1 of the Supplementary Material and data can be accessed using the CWP data platform ().
2.2 Earth observation data and features
2.2.1 Optical imagery and vegetation indices
We rely on MODIS and Sentinel 2 imagery as the basis for all our analysis. A false color composite of dry season imagery reveals large scale irrigation schemes such as those shown in Figure 1 for Zambia. Enhanced Vegetation Index (EVI) time series are used to capture vegetation phenologies to train classifiers on the temporal modality of cropping cycles of the agricultural farm plots.
FIGURE 1
2.2.2 Topographical datasets
Topographic features provide insight on accumulation of moisture and are thus critical to methodology for detecting smallholder irrigation reliant on surface water. We utilize a HydroSHEDS void-filled DEM for elevation data obtained in 2000 by NASA’s Shuttle Radar Topography Mission (SRTM) visualized in Figure 2 (). We also utilize other products derived from DEM such as slope which indicates areas that are unsuitable for agriculture due to the gradient, HAND and MTPI which provide information about local features such as ridges and valley bottoms.
FIGURE 2
Height Above Nearest Drainage (HAND) is a terrain analysis technique that calculates the vertical distance between a point on the landscape and its nearest drainage channel or river. It normalizes elevation data based on the local drainage network, effectively representing how high or low a location is relative to the nearest stream (). Figure 3 shows a visualization of HAND for Uganda (). The SRTM Multi-Scale Topographic Position Index (MTPI) is used to characterize landforms by comparing the elevation of a specific location to the mean elevation of its surrounding neighborhood across multiple spatial scales (). Derived from the Shuttle Radar Topography Mission (SRTM) 30-m Digital Elevation Model (DEM), MTPI is particularly effective for identifying geomorphological features such as valley bottoms, ridge tops, and mid-slope benches.
FIGURE 3
2.2.3 Climatic variables
We analyze feature variables such as precipitation and land surface temperature. In the context of smallholder irrigation detection, Land Surface Temperature (LST) is a critical climatic variable that acts as a thermodynamic proxy for soil moisture and evapotranspiration (ET). While vegetation indices like EVI measure the “greenness” of a plant, LST provides insights into the plant’s actual water stress and the cooling effect of irrigation. MODIS (MOD11A1/A2) provides high temporal resolution (daily) but at a coarse spatial resolution (1 km). CHIRPS provides precipitation data. The WorldClim V1 Bioclim dataset provides 19 bioclimatic variables derived from monthly temperature and precipitation data. It covers the period from 1960 to 1991 and has a resolution of 927.67 m (). This product was preferred over the more recent but coarse ERA5-land product with a resolution of 11,132 m. A scatter plot comparing ERA5-land temperature and precipitation data for 2023 to BIO1 and BIO12 is added to the Supplementary Material. The plot reveals a spatial correlation between the two products despite the temporal mismatch justifying their inclusion in the feature set.
2.2.4 Anthropic variables
Proximity to surface water (such as rivers, streams or permanent wetlands), high resolution settlements, and markets dictate the feasibility of irrigation. Therefore we incorporate anthropic variables that represent the human-driven factors and infrastructure constraints that dictate where irrigation is socio-economically feasible. A description of the datasets used are described in the subsequent paragraph. We threshold the raster layers to obtain a binary mask of the relevant feature and utilize a euclidean-based distance kernel to calculate the distance to the feature. This results in input feature layers such as distance to surface water, distance to night lights, etc., To train the classifiers.
The Global Surface Water (GSW) provides the location and persistence of available surface water (). Meta formerly Facebook in partnership with CIESIN utilized computer vision techniques to map population density in the High-Resolution Settlement Layer (HRSL). This layer provides a proxy for labor availability and urban intensity. Nighttime light (VIIRS) also provides similar insights and therefore is used as part of the feature set. Reliable road access allows for the transport of surplus perishables to urban markets before they spoil. The Global Roads Inventory Project (GRIP) dataset provides recent and consistent global roads dataset (). Protected areas (national parks, wildlife reserves) represent a hard constraint on agricultural expansion. Distance to Protected Areas World Database on Protected Areas (WDPA) is therefore considered in our feature study to determine its impact.
2.2.5 Phenology maps
Phenology maps provide a spatiotemporal characterization of vegetation cycles (). Consistent with our previous study pertaining to dry season smallholder irrigation detection across sub-Saharan Africa, we generate a phenology map of Uganda. We extract temporal end members that represent distinct MODIS 250 m Enhanced Vegetation Index (EVI) time series. Next, we use a MinMaxScaler () to scale the MODIS EVI values to Sentinel 2 to generate a phenology map at 10 m shown in Figure 4. A comparison of the phenology maps from MODIS and Sentinel 2 is provided in the Supplementary Material for validity.
FIGURE 4
The benefits are two-fold. Firstly it allows us to encode the annual temporal signature of the vegetation through mixture modeling of the temporal end members derived from distinct MODIS 250 m Enhanced Vegetation Index (EVI) time series. This reduces the dataset from 12 dimensions corresponding to monthly mean EVI of each pixel for the year to 3 dimensions corresponding to single cropping cycle, dual cropping cycle and evergreen vegetation. The second benefit lies in being able to isolate geographic regions dominated by specific phenologies to assess model performance in those regions.
2.3 Classification, feature selection and validation
Species Distribution Models (SDM) are used to model the habitat suitability for a species by identifying the predictor variables (
). We adapt this methodology to model smallholder irrigation patterns in Sub-Saharan Africa after identifying key predictor variables. This methodology requires information about the occurrence of irrigation which is obtained from survey datasets and various topographic, climatic and anthropic variables described in the previous section. The workflow shown in
Figure 5is as follows:
Collection and preprocessing of survey data
Generate a stack of datasets and variables to use for predictions
Feature selection to identify key predictors
Generation of pseudo-absence data to augment survey data
Model fitting and prediction
Accuracy assessment
FIGURE 5
We generate a stack of input features described earlier (see Section 2 of the Supplementary Material). Using farm plot locations with labels (irrigated, non-irrigated) from the surveys, we extract feature values for each location. This results in a tabular dataset of features values paired with a target binary irrigation class. Next step involves identifying highly correlated feature values to reduce data redundancy, prevent model overfitting and improve computational efficiency. Removing such features increases feature independence enhancing model interpretability by avoiding misleading feature importances.
SHAP (SHapley Additive exPlanations) is a unified framework that utilizes a game-theoretic approach to calculate the marginal contribution of each feature to the model’s output (). In this study, we use SHAP to evaluate the specific predictive importance value of each feature for feature selection. The CatBoost machine learning model is used for irrigation classification which is shown to outperform other gradient boosting frameworks like XGBoost or LightGBM with default hyperparameters ().
The classifier is trained on input features and the feature importance is ranked by SHAP in descending order. The top 20 features are shown in Figure 6. Area of the plots which is obtained from survey data by multiplying the short and long side of the plot is ranked as the most important feature (see Uganda in Supplementary Figure S5 for the full plot). While area is not used for irrigation predictions in our study, it can be incorporated through segmentation of smallholder plots from high-resolution imagery or utilizing existing agricultural field boundary datasets (). A beeswarm plot of the area feature reveals that from the surveyed plots (that do not include the large irrigation schemes), a plot is more likely to be irrigated if it is a smallholder plot. The next two most important features are HAND and SRTMmTPI suggesting that topography is a good predictor of irrigation patterns. Similarly we would expect annual EVI to be a good indicator of irrigation. The model also ranks the latitude of the farm locations as important input features. While we do not use the coordinates for our predictions to enable transfer learning, we hypothesize that this is due to distinct vegetation phenologies across Uganda. Therefore a breakdown of the feature importances across Northern, Central and Southern Uganda is added to the Supplementary Figure S5 of the Supplementary Material.
FIGURE 6
SHAP cannot be used for causal inference. To drive conclusions that extend beyond the model to enable transfer learning, we examine the distributions of the feature values of the irrigated and non-irrigated labels by utilizing a violin plot. Looking at the anthropic variables in Figure 7, we see that these plots are closer to settlements (as shown by DHRSL values) and further from protected areas (as shown by DIPA/DNPA values). DOR1/DOR2 values (distance to major and minor roads) show that irrigated areas are situated closer to infrastructure. While the distributions of these features across irrigated and non-irrigated plots do not provide a meaningful way of segregating the two, we utilize these features to exclude regions where we do not expect irrigation. For example, a buffer of 5 km is applied to the DHRSL feature to reduce the computation cost of identifying irrigated areas. This allows us to develop other similar thresholds if needed to further exclude regions.
FIGURE 7
Due to the class imbalance of more irrigation data points than non-irrigated data points in the survey data, we generate pseudo-absence data points where we expect irrigation to not exist. This is a key step in SDM and can be approached through multiple techniques. One technique involves masking out regions using phenology maps where we do not expect irrigation to exist such as evergreen and barren areas. Another technique used in this methodology generates a 5 km distance buffer layer around the survey data points to sample absence points for non-irrigated areas. It is important to acknowledge the limitation and potential errors of this method that the generation of these pseudo-absence points might involve irrigated areas that were not captured in the survey. We tested two sampling strategies of pseudo-absence points using non-cropped areas from a cropland mask and using a 5 km distance buffer layer. The results can be viewed in the Supplementary Table S2. The 5 km distance buffer layer performed slightly better than the cropland mask suggesting that the results are not an artifact of the negative sampling strategy used. Another technique is to utilize clustering algorithms to generate clusters of presence (irrigated areas) and absence (non-irrigated areas) and sample more pseudo-absence points from the latter cluster.
2.4 Ethics statement
Survey approval was obtained from local authorities along with IRB approval for the 13,840 farmers that were interviewed in Uganda. IRB approval was also obtained for the 95 sites surveyed across Zambia. No uniquely identifiable data of the farmers was collected.
3 Results
3.1 Benchmark performance (Uganda and Zambia)
To assess model performance over a balanced dataset, we first use a subset of the irrigation dataset from the Uganda survey so that the number of irrigated and non-irrigated points are the same. This helps to establish a baseline and facilitate a direct comparison between the data-rich region in Uganda and a data-sparse region in Zambia. To evaluate classifier performance, the dataset consisting of irrigated and non-irrigated farm plot locations are split into 70% training, 30% testing. Next we utilize methods in the Uganda dataset such as pseudo-absence points to augment non-irrigated data points and spatial block cross validation to reduce the impact of spatial autocorrelation between training and testing datasets and improve model generalizability.
Table 1 benchmarks the performance metrics of embeddings vs. feature engineering approach across the two regions. To understand how the embeddings contribute to the final outputs, we compared Linear Regression with a Random Forest model for Uganda. This provides insight into how the structure of the embedding space helps with land cover classification. While a Random Forest model or a neural network might improve the accuracy of the model, it masks whether the information contained in the embedding space is structured and lends to transfer learning. For smallholder irrigation detection in Uganda, the Linear Regression model performs similar to the Random Forest model revealing there is not much non-linear structure to exploit.
TABLE 1
| Region | Method | Algorithm | F1 score | Precision | Recall | Accuracy |
|---|---|---|---|---|---|---|
| Uganda | AEF Embeddings | Linear Regression | 0.711 | 0.703 | 0.718 | 0.716 |
| Random Forest | 0.702 | 0.702 | 0.703 | 0.718 | ||
| Selected features | Random Forest | 0.733 | 0.710 | 0.760 | 0.722 | |
| Zambia | AEF Embeddings | Random Forest | 0.940 | 0.942 | 0.939 | 0.923 |
| Selected features | Random Forest | 0.942 | 0.935 | 0.950 | 0.924 |
Performance metrics of models trained on AEF embeddings and selected features for Uganda (primary evaluation on balanced subset of survey data) and Zambia.
Bold values represent higher performance metrics across the two methods.
In Zambia, the irrigated and non-irrigated polygons are randomly split into the training (70% of polygons) and testing set (30% of polygons). We compare three sampling strategies to sample points from the polygons to study the impact of spatial autocorrelation. The first strategy was to sample all 10 m pixels from the polygons. The other two strategies involved decimation of the nearest one and two neighbors along all four sides of the pixel effectively being sampled at 20m and 30 m scale. This was repeated three times across all sampling strategies for the two methods (AEF embeddings and selected features) for cross-validation. A maximum of 2% increase in average F1 score and accuracy was observed across the sampling strategies with 30 m sampling showing higher performance for both methods. Additionally, the smallholder irrigated plots are of the order of a 10th of a hectare with irrigation methods employed and irrigated crops grown differing across each farmer. Therefore, the marginal F1 change observed from the pixel-decimation robustness check can therefore be used to conclude that a 30 m sampling strategy reasonably reduces spatial autocorrelation effects. Table 1 reports the average performance metrics for 30 m sampling across three different random splits of polygons into training and testing sets. Detailed performance metrics for all sampling strategies are added to the Supplementary Material.
Figure 8 shows the output of a Random Forest model trained using AEF embeddings and selected features for Zambia. We see that the outputs based on embeddings overpredict irrigation across Central Zambia away from settlement. These areas are masked using distance to high resolution settlement in the feature engineering approach demonstrating the benefits of interpretable features and extension of conclusions drawn from the feature study in Uganda.
FIGURE 8
Next using a spatial block cross-validation approach using the entire Uganda survey data, we train 5 Random Forest models using AEF embeddings and a selection of input features. The spatial blocking ensures cross-validation across the distinct phenologies in Uganda shown in Figure 4 earlier. Looking across the performance metrics in Table 2, we see that the AEF embeddings outperform the selected features for Uganda. The irrigation predictions using these models is shown in Figure 9. Overall similar patterns of irrigation are observed across the two outputs with higher irrigation predictions across South-Western Uganda and the crescent along Lake Victoria. The AEF embeddings based model output predicts more irrigation than the selected features based output especially along the cattle corridor. Another noticeable feature of the latter output is the reliance on topography (from HAND) to predict irrigation patterns in Central Uganda.
TABLE 2
| Method | F1 | Precision | Recall | AUC | Accuracy |
|---|---|---|---|---|---|
| AEF Embeddings | 0.824 | 0.777 | 0.876 | 0.886 | 0.886 |
| Selected Features | 0.809 | 0.796 | 0.822 | 0.863 | 0.788 |
Average performance metrics of 5 Random Forest models trained on AEF embeddings and selected features for Uganda using spatial block cross validation approach. Higher performance metrics across the two methods are represented in bold.
Bold values represent higher performance metrics across the two methods.
FIGURE 9
3.2 Feature interpretation and variable ablation
To understand where smallholder farmers are irrigating in Uganda, different remote sensing features for the locations identified from the survey as irrigated and non-irrigated fields are compared through a violin plot. This allows us to see if the distribution for each feature varies significantly for irrigated vs. non-irrigated croplands and to judge if the feature would be useful for training a classifier. Feature distribution for annual EVI (Enhanced Vegetation Index) show higher values for irrigated croplands as one would expect. Irrigated areas are found at lower lying areas where water accumulates and this is evident by looking at the distribution of the multi-scale topographic index which shows the mean elevation of a region. Another method of judging the importance of a set of variables can be achieved by omitting them from the model. Figure 10 shows an example of this where the omission of anthropic variables results in the model predicting irrigation in the cattle corridor in South Western Uganda.
FIGURE 10
In Zambia, smallholder irrigation is heavily reliant on the riverine areas (dambos) which can be identified from Digital Elevation Models (DEM). By examining Sentinel 2 dry season imagery, we can identify irrigated croplands because they remain greener than surrounding vegetation. However, this greenness alone does not distinguish irrigated croplands from riparian vegetation. This can be observed by looking at the monthly Sentinel 2 EVI time series of irrigated, non-irrigated and forests/riparian areas in Figure 11. In some cases, the EVI patterns of riparian areas are very similar to those of irrigated croplands, making those areas difficult to distinguish using EVI alone.
FIGURE 11
3.3 Transferability
To assess transferability, we perform a quantitative out of region validation. This is done by visual labeling of polygons around other REMP sites. While this method is not as definitive as ground-collected labels used for training, high quality labels can be collected by examining high-resolution imagery. Different historic dates are examined using Google Earth Pro imagery to find traits similar to those observed while labeling polygons around ground truth. If an area appears to be cropped near dambos which is easily differentiable from uncropped riparian vegetation due to crop-spacing and hedges on the outskirts used to deter livestock, it is labeled as irrigated. This process resulted in 881 visually labeled irrigated polygons. We then utilize a cropland mask to sample non-irrigated points outside the cropped areas.
The performance of the two approaches assessed on this dataset is shown in Table 3. On average, the Random Forest model trained on AEF embeddings has a higher F1 score and Recall whereas the model based on feature engineering has a higher overall accuracy and Precision. Qualitatively we observe in Figure 8 that the model trained on embeddings overpredicts irrigation in Central Zambia.
TABLE 3
| Method | Sampling | Run | F1 score | Precision | Recall | Accuracy |
|---|---|---|---|---|---|---|
| AEF Embeddings | 30 m | 1 | 0.721 | 0.652 | 0.806 | 0.704 |
| 2 | 0.738 | 0.616 | 0.919 | 0.691 | ||
| 3 | 0.726 | 0.601 | 0.918 | 0.673 | ||
| Average | 0.728 | 0.623 | 0.881 | 0.689 | ||
| Selected features | 30 m | 1 | 0.608 | 0.673 | 0.555 | 0.664 |
| 2 | 0.731 | 0.693 | 0.774 | 0.732 | ||
| 3 | 0.756 | 0.619 | 0.973 | 0.740 | ||
| Average | 0.700 | 0.661 | 0.767 | 0.712 |
Performance metrics for different model runs for transferability in Zambia. Higher performance metrics across the two methods are represented in bold.
Bold values represent higher performance metrics across the two methods.
4 Discussion
Our findings demonstrate a clear trade-off between model efficiency and interpretability. In regions like Zambia where labels are scarce, AEF embeddings perform at par with the traditional feature engineering approach. This suggests that foundation models offer a great alternative for rapid land cover classification in areas with limited ground-truth data. This offers a practical pathway in the absence of a well-defined feature set constructed through the analysis of labor-intensive ground-truth survey data. In the data-rich context of Uganda, the embeddings based classifier outperformed the feature engineering approach on most performance metrics.
While embedding models are highly effective for classification, they function largely as “black boxes.” Furthermore geospatial foundational models suffer from lack of standardization in evaluations, testing and training protocols and release of pretrained weights (). This lack of transparency presents challenges for researchers who require auditable results for policy or causal inference. Conversely, the feature engineering approach provides a clear, operational workflow. Analysis through SHAP and feature distribution plots revealed that topography and vegetation phenology from EVI time series remain the most potent indicators of smallholder irrigation. These variables successfully capture the preference of farmers to irrigate in low-lying areas where surface water naturally accumulates. Such interpretability allows for the development of specific thresholds to mask out regions for inference reducing the computational cost. It also provides insights into the socio-economic drivers of irrigation, such as proximity to surface water and infrastructure.
Certain methodological choices made in this study might affect results when applied using different techniques or to other regions. This includes the choice of pseudo-absence sampling strategy. While a comparison of using a 5 km buffer around irrigated samples performed similar to utilizing a cropland mask (see Supplementary Table S2), other strategies such as using phenology maps for negative sampling might yield different results. Similarly another caveat is the transferability which was assessed using visually labeled irrigated polygons and non-irrigated points sampled from a cropland mask. Finally, the choice of embeddings from the geospatial foundation models mentioned in the introduction can impact performance. These foundational models offer different embedding dimensions such as 64 for AlphaEarth Foundations, 128 for OlmoEarth, 256 for TESSERA and 1024 dimensions for Prithvi-EO-2.0. However, a PCA variance analysis revealed that most embedding dimensions are redundant suggesting that the extra capacity adds noise rather than signal ().
Beyond agricultural monitoring, the ability to map irrigation clusters has direct implications for energy infrastructure. Smallholder irrigation often represents a concentrated captive power demand that can justify the deployment of mini-grids or the extension of the national grid. By accurately detecting these areas, developers can better estimate energy and water requirements, facilitating the transition from manual labor or diesel generators to more sustainable, electricity-based pumping solutions.
5 Conclusion
The identification of smallholder irrigation is fundamental to understanding the productive uses of energy in Sub-Saharan Africa. This study provides a comprehensive benchmarking of two distinct remote sensing paradigms: the geospatial foundation model embeddings from AlphaEarth Foundations (AEF) and a traditional feature engineering approach incorporating topographic, climatic, and anthropic variables. By utilizing survey data from Uganda and Zambia, we evaluated how these methods perform across varying levels of data density and climatic complexity.
This research highlights that geospatial foundation models like AlphaEarth Foundations provide quick classification that outperform or remain on par with feature engineering methods for smallholder irrigation detection in Sub-Saharan Africa. However the feature engineering approach remains indispensable for generating interpretable, auditable, and transferable insights. Future efforts should focus on hybrid workflows that leverage the predictive power of embeddings while maintaining the transparency necessary for impactful policy decisions and energy access planning.
Statements
Data availability statement
The Google Earth Engine scripts and the Zambia training data used for the results in this paper are available on Github at github.com/hssiddiqui/Irrigation-Detection-Uganda-Zambia. The Uganda Geodata can be requested via Redivis (). A district level summary of the Uganda survey data is available on the CWP data platform.
Ethics statement
The studies involving humans were approved by Institutional Review Board (IRB). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.
Author contributions
HS: Investigation, Writing – original draft, Formal Analysis, Data curation, Visualization, Validation, Conceptualization, Writing – review and editing. CS: Supervision, Writing – review and editing. VM: Writing – review and editing, Funding acquisition, Supervision.
Funding
The author(s) declared that financial support was received for this work and/or its publication. Partial support for this effort was provided by the National Science Foundation (INFEWS Award Number 1639214), Columbia World Projects, Rockefeller Foundation (eGuide Grant 2018POW004), OPML United Kingdom (DFID) and Technoserve (BMGF).
Acknowledgments
The authors acknowledge Markus Walsh for discussion about relevant feature sets and ensemble methods. The authors also acknowledge the role of Edwin Adkins and Aftab Zindani in survey design and training the survey teams to collect data without which this study would not have been possible.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was used in the creation of this manuscript. Generative AI (Gemini 2.5 Series) was used in the creation of this manuscript. The tool was used exclusively to write Google Earth Engine code which was modified later by the authors and enhance the readability of the manuscript. The AI system was not used to generate scientific ideas, perform data analysis, or interpret results. All study design, data processing, analyses, and conclusions are entirely the work of the authors, who reviewed and verified all manuscript content.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsen.2026.1895172/full#supplementary-material
References
1
BjornlundV.BjornlundH.Van RooyenA. (2022). Why food insecurity persists in Sub-Saharan Africa: a review of existing evidence. Food Secur.14 (4), 845–864. 10.1007/s12571-022-01256-1
2
BoastR. (1990). Dambos: a review. Prog. Phys. Geogr. Earth Environ.14 (2), 153–177. 10.1177/030913339001400201
3
BrownC. F.KazmierskiM. R.PasquarellaV. J.RucklidgeW. J.SamsikovaM.ZhangC.et al (2025). AlphaEarth foundations: an embedding field model for accurate and efficient global mapping from sparse label data. arXiv. 10.48550/ARXIV.2507.22291
4
BullockA. (1992). Dambo hydrology in southern Africa—Review and reassessment. J. Hydrology134 (1–4), 373–396. 10.1016/0022-1694(92)90043-U
5
BurkeM.DriscollA.LobellD. B.ErmonS. (2021). Using satellite imagery to understand and promote sustainable development. Science371 (6535), eabe8628. 10.1126/science.abe8628
6
Columbia World Projects Energy for Productive Use (2026). Available online at: https://qsel.columbia.edu/cwp-epu-data-platform/.
7
ConlonT.WuY.SmallC.SiddiquiH.AdkinsE.ModiV.et al (2020). A novel method of irrigation detection and estimation of the effects of productive electricity demands on energy System planning. AGU Fall Meet. Abstr., GC034–GC038.
8
ConlonT.SmallC.ModiV. (2022). A multiscale spatiotemporal approach for smallholder irrigation detection. Front. Remote Sens.3, 871942. 10.3389/frsen.2022.871942
9
CorleyI.LehmannN.RobinsonC.TsengG.FullerA.AlemohammadH.et al (2026). No one knows the state of the art in geospatial foundation models. arXiv. 10.48550/ARXIV.2605.12678
10
CregoR. D.StabachJ. A.ConnetteG. (2022). “Implementation of species distribution models in Google Earth engine,” Divers. Distrib. Edited by LeroyB., 28(5), pp. 904–916. 10.1111/ddi.13491
11
DonchytsG.WinsemiusH.SchellekensJ.EricksonT.GaoH.SavenijeH. G.et al (2016). “Global 30m height above the nearest drainage. 10.13140/RG.2.1.3956.8880
12
DunningC. M.BlackE. C. L.AllanR. P. (2016). The onset and cessation of seasonal rainfall over Africa. J. Geophys. Res. Atmos.121 (19). 10.1002/2016jd025428
13
FengZ.AtzbergerC.JafferS.KnezevicJ.SormunenS.YoungR. (2025). TESSERA: temporal embeddings of surface spectra for Earth representation and analysis. arXiv. 10.48550/ARXIV.2506.20380
14
GillerK. E.DelauneT.SilvaJ. V.DescheemaekerK.van de VenG.SchutA. G.et al (2021). The future of farming: who will produce our food?Food Secur.13 (5), 1073–1099. 10.1007/s12571-021-01184-6
15
HaanyikaC. M. (2008). Rural electrification in Zambia: a policy and institutional analysis. Energy Policy36 (3), 1044–1058. 10.1016/j.enpol.2007.10.031
16
HerzogH.BastaniF.ZhangY.TsengG.RedmonJ.SablonH. (2025). OlmoEarth: stable latent image modeling for multimodal Earth observation. arXiv. 10.48550/ARXIV.2511.13655
17
HijmansR. J.CameronS. E.ParraJ. L.JonesP. G.JarvisA. (2005). Very high resolution interpolated climate surfaces for global land areas. Int. J. Climatol.25 (15), 1965–1978. 10.1002/joc.1276
18
KernerH.ChaudhariS.GhoshA.RobinsonC.AhmadA.ChoiE. (2024). Fields of the world: a machine learning benchmark dataset for global agricultural field boundary segmentation. arXiv. 10.48550/ARXIV.2409.16252
19
KocamanA. S.OzyorukE.TanejaS.ModiV. (2020). A stochastic framework to evaluate the impact of agricultural load flexibility on the sizing of renewable energy systems. Renew. Energy152, 1067–1078. 10.1016/j.renene.2020.01.129
20
LehnerB.VerdinK.JarvisA. (2008). New global hydrography derived from spaceborne elevation data. Eos, Trans. Am. Geophys. Union89 (10), 93–94. 10.1029/2008EO100001
21
LundbergS. M.LeeS.-I. (2017). A Unified Approach to Interpreting Model Predictions. 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA. Curran Associates, Inc.Available online at: http://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf (Accessed August 29, 2025).
22
MabhaudhiT.MpandeliS.NhamoL.ChimonyoV.NhemachenaC.SenzanjeA.et al (2018). Prospects for improving irrigated agriculture in Southern Africa: linking water, energy and food. Water10 (12), 1881. 10.3390/w10121881
23
MeijerJ. R.HuijbregtsM. A. J.SchottenK. C. G. J.SchipperA. M. (2018). Global patterns of current and future road infrastructure. Environ. Res. Lett.13 (6), 064006. 10.1088/1748-9326/aabd42
24
NobreA. D.CuartasL.HodnettM.RennóC.RodriguesG.SilveiraA.et al (2011). Height above the Nearest drainage – a hydrologically relevant new terrain model. J. Hydrology404 (1–2), 13–29. 10.1016/j.jhydrol.2011.03.051
25
OzdoganM.YangY.AllezG.CervantesC. (2010). Remote sensing of irrigated agriculture: opportunities and challenges. Remote Sens.2 (9), 2274–2304. 10.3390/rs2092274
26
PekelJ.-F.CottamA.GorelickN.BelwardA. S. (2016). High-resolution mapping of global surface water and its long-term changes. Nature540 (7633), 418–422. 10.1038/nature20584
27
ProkhorenkovaL.GusevG.VorobevA.DorogushA. V.GulinA. (2017). CatBoost: unbiased boosting with categorical features. arXiv. 10.48550/ARXIV.1706.09516
28
RobinsonC.CorleyI. (2026). Compressing Earth Embeddings. Available online at: https://geospatialml.com/posts/compressing-earth-embeddings/.
29
SiddiquiH.ModiV. (2024). “Monitoring the impacts of disruption events on agriculture through irrigation detection with remote sensing,” in 2024 IEEE Global Humanitarian Technology Conference (GHTC). 2024 IEEE Global Humanitarian Technology Conference (GHTC) (Radnor, PA, USA: IEEE), 286–291. 10.1109/GHTC62424.2024.10771556
30
SiddiquiH.SmallC.ModiV. (2025). Operationalizing remote sensing methods for smallholder dry season irrigation detection in Sub-Saharan Africa. Front. Remote Sens.6, 1661528. 10.3389/frsen.2025.1661528
31
SmallC. (2012). Spatiotemporal dimensionality and Time-space characterization of multitemporal imagery. Remote Sens. Environ.124, 793–809. 10.1016/j.rse.2012.05.031
32
SzwarcmanD.RoyS.FraccaroP.GíslasonP. E.BlumenstielB.GhosalR. (2024). Prithvi-EO-2.0: a versatile multi-temporal foundation model for Earth observation applications. arXiv. 10.48550/ARXIV.2412.02732
33
The Quadracci Sustainable Engineering Lab (2026). Uganda geodata. Redivis. 10.71778/V2DW-7A53
34
TheobaldD. M.Harrison-AtlasD.MonahanW. B.AlbanoC. M. (2015). “Ecologically-Relevant maps of landforms and physiographic diversity for climate adaptation planning,”. Editor CarmelY., 10, e0143619. 10.1371/journal.pone.0143619
35
TurnerB. (1986). The importance of dambos in African agriculture. Land Use Policy3 (4), 343–347. 10.1016/0264-8377(86)90030-X
36
Uganda Observed Climatology of Precipitation (no date) Climate Change Knowledge Portal (2026). Available online at: https://climateknowledgeportal.worldbank.org/country/uganda/climate-data-historical (Accessed: May 19, 2022).
37
WanyamaJ.SseganeH.KisekkaI.KomakechA. J.BanaddaN.ZziwaA.et al (2017). Irrigation development in Uganda: constraints, lessons learned, and future perspectives. J. Irrigation Drainage Eng.143 (5), 04017003. 10.1061/(asce)ir.1943-4774.0001159
Summary
Keywords
alphaearth foundations embeddings, google earth engine (GEE), irrigation, phenology, SHAP (shapley additive exPlanations), Uganda, Zambia
Citation
Siddiqui H, Small C and Modi V (2026) From foundational models to interpretability: a comparative study of embeddings and feature engineering for smallholder irrigation detection in Sub-Saharan Africa. Front. Remote Sens. 7:1895172. doi: 10.3389/frsen.2026.1895172
Received
29 May 2026
Revised
17 July 2026
Accepted
31 July 2026
Published
28 August 2026
Volume
7 - 2026
Edited by
Qian Yu, University of Massachusetts Amherst, United States
Reviewed by
Bereket Tsehaye Haile, University of Pretoria, South Africa
Zolo Kiala, International Water Management Institute, South Africa
Updates
Copyright
© 2026 Siddiqui, Small and Modi.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Hasan Siddiqui, hss2152@columbia.edu
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.