Bananas and plantains rank among the world’s top 10 food crops (Kumar et al., 2015). In East Africa, bananas serve as a staple food for over seven million people and are a primary source of income for millions of smallholder farmers (Gold et al., 2002). Over the past few decades, several significant banana diseases have emerged and spread extensively across different regions, posing a risk to food security and jeopardizing the livelihoods of smallholder farmers (Tatineni and Hein, 2023).

Banana bunchy top disease, caused by the banana bunchy top virus (BBTV: genus Babuvirus, family Nanoviridae), is the most destructive viral disease affecting bananas globally, posing a significant threat to smallholder banana cultivation. The virus is primarily transmitted by the aphid vector Pentalonia nigronervosa and further disseminated using infected propagation materials (Magee, 1927, 1940). Banana Bunchy Top Virus infection can result in yield losses of up to 100% (Okonya et al., 2019).

Over the past decade, BBTV has spread across sub-Saharan Africa, with confirmed cases in the Democratic Republic of Congo, Angola, Cameroon, Gabon, Malawi (Kumar et al., 2011), Nigeria (Adegbola et al., 2013), Benin (Lokossou et al., 2012), Togo (Kolombia et al., 2021), Uganda (Ocimati et al., 2021, 2024), and Tanzania (Shimwela et al., 2022). Recent findings confirm that BBTV has been established in Tanzania, with BBTV being found in 10 regions (Mahuku and Kumar, 2023). Urgent interventions are needed to halt the spread of the virus throughout the country.

Mathematical models of epidemics offer valuable insights into the mechanisms driving disease spread and enable comparison of different management strategies for viral spread. There are only a few large-scale spatially explicit models of plant disease spread, including sudden oak death (Cunniffe et al., 2016), Xylella fastidiosa in olive trees (White et al., 2017), citrus Huanglongbing disease (Mastin et al., 2020; Nguyen et al., 2023), cassava brown streak disease (Godding et al., 2023), and brown rot in peach (Radici et al., 2024). One important limiting factor is the absence of host maps upon which to model pathogen spread in the target countries.

We introduced a novel approach for generating banana cultivation maps using high-resolution remote sensing data. Our method involves identifying a sample of locations with banana cultivation through photo interpretation of high-resolution satellite imagery. We then created binary maps indicating the presence or absence of banana by utilizing high-resolution data on vegetation indices, canopy height, and built-up areas. Remote sensing and production data at the regional level were used to construct a banana production map of Tanzania. Recently, high-resolution RGB and multispectral aerial imagery from an unmanned aerial vehicle (UAV) have been deployed to identify bananas in smallholder farming systems in Ogun State, Nigeria (Alabi et al., 2022). We tested our proposed method against predictions based on UAV imagery. Finally, we used the host map for Tanzania together with limited published survey data from 2020 and 2023 to parameterize a model for the spread and transmission of BBTV at country-wide scales.

This study presents the first comprehensive stochastic model that outlines methods for developing data-driven models to generate banana production maps, and process-based models to evaluate the potential spread of BBTV on a landscape scale. Few studies have been conducted to model the spread of BBTV at a plantation or field scale (Allen and Barnier, 1977; Allen, 1987; Smith et al., 1998; Varghese et al., 2020). Plantation-scale models provide tools for optimizing control of commercial farms. Recently, regression-based methods have been applied to identify areas that have high environmental suitability for BBTV establishment, such as Uganda (Ocimati et al., 2024) and the entire African continent (Bouwmeester et al., 2023). However, maps showing suitability for BBTV cannot be used to inform smallholder farmers of the best practices for mitigation facing the threat of the imminent spread of BBTV.

Our methodology for mapping banana cultivation showed good agreement with the results based on machine learning combined with high-resolution RGB and multispectral aerial imagery conducted in Ogun State in Nigeria (Alabi et al., 2022). Besides banana and plantain, a variety of crops are cultivated in the state, including cocoa, oil palm, oranges, maize, cassava, cowpea, and vegetables, with farmers predominantly practicing intercropping (Alabi et al., 2022). We found a high degree of correlation between the estimated banana area based on the approach introduced in this study and the estimates from ML and UAV images (ρ = 0.915–0.926). Pixel-based classifications and ML models require the manual delineation and annotation of thousands of reference points (Gomez Selvaraj et al., 2020). For example, at the Olokuta site, 223,578 georeferenced polygons were derived for different classes (Alabi et al., 2022). Therefore, because of the amount of effort required to obtain and process UAV-derived imagery, it is not possible to scale this method to the national level. The proposed workflow for map construction can be deployed at the country level, as demonstrated in the current study. We used publicly available data to derive a high-resolution banana production map. In contrast, there is no repository of UAV-derived imagery.

Our methods require the distribution of canopy height and NDVI values typical for bananas as input. We used photo-interpretation of high-resolution satellite imagery from Google Earth Pro to sample locations and derive these distributions. The range for canopy height was between 2 m and 6 m, and the annual median NDVI was in the range of ∈ [0.55,0.88]. Many factors can influence banana plant height, such as stage of growth (Tixier et al., 2004), cultivar (Daniells and O’Farrell, 1988), planting density, and irrigation (zum Felde et al., 2016). Plant photosynthetic activity also has a complex dependence on seasonality and environmental variables (Brown and de Beurs, 2008). The wide ranges of canopy height and annual median NDVI distributions that we obtained reflect this variability. However, the methodology allowed us to derive a probabilistic host distribution, which to the best of our knowledge, has not been performed before.

Our methodology has several limitations. First, it requires a large amount of data ( Table 1 ). Another aspect is that the algorithm performed well in areas with high banana presence (i.e., in the Kagera region), but overestimated banana density in areas where banana was grown sparsely or where intercropping was common. The performance of our method depended on the type of intercropping used. When bananas are grown in association with annual food crops, such as maize, rice, and cassava, we expect reasonable efficiency (as seen in Figure 2B ) The efficiency will be low when bananas are combined with perennial trees. For example, oil palm trees have a morphology similar to that of bananas. In Tanzania, the Kigoma region has the largest harvested area of oil palm (4,726 ha), followed by Mbeya (1,614 ha), whereas the least harvested area of oil palm is in Kagera (2 ha) (National Sample Census of Agriculture, 2019/2020). There is also a diverse variety of permanent tree crops cultivated across various regions of Tanzania (Kilawe, 2024), with fruit trees present on almost every farm (Delobel et al., 1991). To overcome this limitation, we normalized the total area planted with bananas in each region using data from the Tanzania National Sample Census of Agriculture 2019–2020 (National Sample Census of Agriculture, 2019/2020). However, the use of national data to address overestimation in areas with low production may face challenges in regions where the data quality is poor or entirely lacking within a country or production landscape. Additional information, such as the locations of oil palms, can be utilized to mask such areas, as in the case of Ogun State.

We introduced a novel parameter estimation method by proposing a score metric that assesses the similarity of simulated outbreaks to survey data at the district level. To parameterize the model for BBTV spread in Tanzania, we used the survey results from 2020 (Shimwela et al., 2022) and 2023 (Mahuku and Kumar, 2023) aggregated at the district level. To account for the variability in district areas, we normalized the contribution of each district by the total number of grid cells within a district. This score also accounts for stochasticity in the model outputs. The range of permissible values of the score metric lie between zero and one, where one corresponds to simulations reproducing survey results exactly at the district level. For the model simulations using the fitted parameters, we obtained M ₂₀₂₀ = 0.985 and M ₂₀₂₃ = 0.851. Another novel aspect of the ABC rejection algorithm we used is maximizing the score metric instead of minimizing a summary statistic.

A key assumption in demonstrating the approach to modeling BBTD spread in Tanzania was that the disease was confined to the districts in which it had been reported during the surveys. Another key aspect of this approach is model validation. Only a few studies have validated the spatial spread of crop pathogens and pests on a national scale, which requires high-resolution spatial and temporal surveillance data (Nguyen et al., 2023; Godding et al., 2023; Retkute et al., 2024). These two aspects highlight the critical role of surveys, as data on the presence and absence of the disease at various epidemic stages are essential not only for disease management, but also for accurate prediction. Regular proactive surveillance plays a key role in providing updated data on the current status of the epidemic spread for use in initiating model predictions of future spread. Second, updated surveillance provides an invaluable resource from which to update model parameterizations and validate models by comparing model predictions with ground-based observations.

Our results have important implications for BBTD management. In Tanzania, bananas are a vital food and commercial crop and serve as a major source of raw materials for the beverage and handicraft industries (Luzi-Kihupi et al., 2015), with almost two million households involved in banana production activities (National Sample Census of Agriculture, 2019/2020). Banana is an important primary staple crop, with annual banana consumption reaching 500 kg–1,500 kg per head in southeastern Kilimanjaro (Yamaguchi and Araki, 2004). Simulated patterns of BBTV transmission in Tanzania emphasize the key areas requiring focused attention for combating BBTV through surveillance and preventive measures to control its spread, i.e., Kagera, Kilimanjaro, Dar es Salaam, and Mbeya. Future work will focus on using the BBTV transmission model to evaluate control options, including roguing, impact of clean seed production networks, and clean seed deployment.