Sub1 and non-Sub1 rice plot locations overlaid on 2021 EVI classification in northern Bangladesh. Sentinel-2 imagery was used to generate Enhanced Vegetation Index (EVI) classifications of rice and fallow land across flood-prone districts
This evaluation explored whether satellite data could detect the presence of Sub1 rice, a flood-tolerant variety developed to survive temporary submergence. The study focused on low-lying rice-producing areas in Bangladesh, where researchers from DevGlobal, AgriTech, and IRRI combined ground survey data with optical and radar imagery to assess Sub1 adoption and performance. The team tested whether flood patterns and vegetation trends could help differentiate Sub1 from conventional rice, and examined how environmental conditions shaped what Earth observation could (and couldn’t) reveal.
Sub1 rice only reveals its special trait under very specific flood conditions. This creates a unique detection challenge. The variety is designed to survive five to fourteen days of submergence. If flooding is shorter, it behaves like any other rice. If longer, all varieties fail.
To identify Sub1 adoption, the team relied on multiple data sources. Sentinel-1 synthetic aperture radar was used to detect flooded fields. Sentinel-2 optical imagery supported vegetation analysis. Long-term vegetation trends were tracked using MODIS EVI and NDVI indices. Most critically, ground surveys conducted by IRRI in 2021 and 2022 provided a reference dataset to validate satellite-based classifications.
But there was a catch. In 2022, floods were minimal. Without submergence, Sub1 rice looked identical to conventional varieties. As David Bergvinson from DevGlobal put it, “We were hoping to validate the 2021 model with 2022 data, but the lack of flooding in 2022 created a major obstacle.”
Sub1 rice was extremely rare in the training data. Among the 477 plots surveyed by IRRI in 2021, only 13 were Sub1. That small sample made it hard for machine learning models to learn reliable patterns. The researchers tested three different classification models using combinations of radar and vegetation indices. EVI-based models performed better than those using NDVI, but the overall accuracy remained low.
The team tried to improve the models by focusing on Kurigram District, where flooding was significant and more Sub1 plots had been verified. They also removed mislabeled plots from the training data and consulted IRRI breeders to ensure varieties were correctly classified. These steps helped, but the limited data and narrow conditions meant the models still struggled to distinguish Sub1 rice from other types.
This case highlights several technical and practical barriers to using EO for tracking submergence-tolerant crops.
First, the expression of Sub1’s key trait is conditional. If flooding does not occur within a narrow window, Sub1 is indistinguishable in satellite imagery. This “Goldilocks” problem means that timing and environmental conditions are critical.
Second, validation efforts were hindered by year-to-year variability. The 2022 season had too little flooding, preventing comparison with 2021 and limiting the ability to test whether the model held up across seasons.
Third, even subtle mislabeling in ground-truth data can ripple through the analysis. Early in the process, salinity-tolerant rice varieties were mistakenly labeled as Sub1. Once these were corrected in consultation with breeders, model performance improved. This reinforces the importance of careful survey design and breeder involvement in EO-driven evaluations.
Despite the modeling challenges, the satellite data offered real value. Sentinel-1 successfully mapped areas that experienced flooding. This helped the researchers target the flooded areas that would be critical for understanding where Sub1 rice was most likely to be planted and to focus survey efforts where the trait could be observed. In terms of classification, EVI provided a clearer picture of potential Sub1 adoption than NDVI, which tended to scatter predictions across the landscape in unrealistic patterns.
One key takeaway is the importance of aligning EO analysis with field data collection. Future evaluations should prioritize denser, georeferenced sampling and stronger coordination with breeders to ensure data quality. The team is also exploring whether training data from other regions, like Odisha in India, could support better model performance.
As Bergvinson noted, this is just one step in a broader inquiry. “How might we extend this learning to other stress-tolerant traits, drought tolerance, salinity tolerance, or regenerative agriculture practices?”
This Bangladesh case reflects both the promise and complexity of using EO to evaluate climate-resilient agriculture. The ability to monitor flood conditions and identify high-risk areas in near real-time can help direct fieldwork and resources. But detecting the presence and impact of traits like Sub1 requires more than just pixels. It also demands a careful match between available data sources, ground conditions, and evaluation design.
