Assessment of a new GeoAI foundation model for flood inundation mapping

Classified as probably not a substantial claim.

Deep learning models have demonstrated the capability to recognize distinctive flood patterns and subtle flood boundaries, even in regions under complex and rapidly changing conditions.

“Through data-driven analysis, deep learning models have demonstrated the capability to recognize distinctive flood patterns and subtle flood boundaries, even in regions under complex and rapidly changing conditions.”

• cites [13] 13. Chia-Yu Hsu and Wenwen Li. 2023. Explainable GeoAI: can saliency maps help interpret artificial intelligence's learning process? An empirical study on natural feature detection. International Journal of Geographical Information Science 37, 5 (2023), 963-987. sourceNot checked — no full text of the source
• cites [14] 14. Chia-Yu Hsu, Wenwen Li, and Sizhe Wang. 2021. Knowledge-driven GeoAI: Integrating spatial knowledge into multi-scale deep learning for Mars Crater detection. Remote Sensing 13, 11 (2021), 2116. sourcePartly supported

  The Introduction describes GeoAI’s potential for automated detection of natural terrain features, and §4.5 reports that the model detected Mars craters of different sizes. Those passages support the general point that deep learning can recognize terrain features, but they do not establish its ability to identify flood patterns or subtle flood boundaries under rapidly changing conditions. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [15] 15. Yingjie Hu, Song Gao, Dalton Lunga, Wenwen Li, Shawn Newsam, and Budhendra Bhaduri. 2019. GeoAI at ACM SIGSPATIAL: progress, challenges, and future directions. Sigspatial Special 11, 2 (2019), 5-15. sourceNot checked — no full text of the source
• cites [23] 23. Wenwen Li, Chia-Yu Hsu, and Maosheng Hu. 2021. Tobler's First Law in GeoAI: A spatially explicit deep learning model for terrain feature detection under weak supervision. Annals of the American Association of Geographers 111, 7 (2021), 1887-1905.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [24] 24. Wenwen Li, Chia-Yu Hsu, Sizhe Wang, Chandi Witharana, and Anna Liljedahl. 2022. Real-time GeoAI for high-resolution mapping and segmentation of arctic permafrost features: the case of ice-wedge polygons. In Proceedings of the 5th ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery. 62-65. sourceNot checked — no full text of the source
• cites [35] 35. Sizhe Wang and Wenwen Li. 2021. GeoAI in terrain analysis: Enabling multi- source deep learning and data fusion for natural feature detection. Computers, Environment and Urban Systems 90 (2021), 101715. sourceNot checked — no full text of the source

May combine two claims.

The billion-scale GeoAI foundation model is not open-sourced, so its adaptability in different domain tasks is difficult to assess.

“However, the model is not open-sourced; therefore, its adaptability in different domain tasks is difficult to assess.”

• cites [6] 6. Keumgang Cha, Junghoon Seo, and Taekyung Lee. 2023. A billion-scale founda- tion model for remote sensing images. arXiv preprint arXiv:2304.05215 (2023). sourceCould not verify

  The Methods (p. 6) describe adapting the model for rotated object detection and semantic segmentation, and the paper reports evaluations on those tasks. But the passages found when searching for open-source code, released weights, and model availability do not say whether this model was open-sourced or whether its availability makes assessment across other domain tasks difficult. (LLM, conf=0.72)

  Experimental — accuracy not yet measured

May combine two claims.

U-Net is a classic model for semantic segmentation and one of the most popular models adopted for flood inundation mapping.

“The U-Net model is selected for performance comparison with Prithvi because it is a classic model for semantic segmentation [33] and one of the most popular models adopted for flood inundation mapping [19].”

• cites [33] 33. Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer, 234-241. sourceSupported
• cites [19] 19. Goutam Konapala, Sujay V Kumar, and Shahryar Khalique Ahmad. 2021. Ex- ploring Sentinel-1 and Sentinel-2 diversity for flood inundation mapping using deep learning. ISPRS Journal of Photogrammetry and Remote Sensing 180 (2021), 163-173. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Transformer-based architecture has shown stronger performance than CNN-based models due to its capability in capturing data dependencies over long ranges and reducing inductive bias.

“Recently, the transformer-based architecture has shown even stronger performance than CNN-based models due to its capability in capturing data dependencies over long ranges and the ability to reduce inductive bias [22].”

• cites [22] 22. Wenwen Li and Chia-Yu Hsu. 2022. GeoAI for large-scale image analysis and machine vision: Recent progress of artificial intelligence in geography. ISPRS International Journal of Geo-Information 11, 7 (2022), 385. sourcePartly supported

  Section 4.5 (page unavailable) says transformers can use a more flexible receptive field than CNNs, and that Vision Transformers surpass CNNs in image classification. It also describes self-attention across all word pairs, consistent with capturing long-range dependencies. The claim is broader than the source: the paper says CNNs still perform better on object detection and semantic segmentation, and it does not attribute transformer performance to reduced inductive bias. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

May combine two claims.

Segformer-B0 contains 3.7M trainable parameters, much smaller than the other models.

“Segfomer-B0 only contains 3.7M parameters (Table 2), which is much smaller than the other models”

May combine two claims. The quote only partly states this claim.

On the unseen Bolivia data, Prithvi yields the highest mIoU and prediction accuracy (mAcc) among all comparative models, outperforming other models in segmenting images in unseen regions.

“Table 3's results show that Prithvi yields the highest mIoU and prediction accuracy (mAcc) among all comparative models.”

• Also stated in: § 4 RESULTS AND ANALYSIS · plain

  “It can be seen that Prithvi predicted better results than the other models, especially in the highlighted image regions (red boxes).”

• Also stated in: § 5 CONCLUSION · plain

  “the foundation model Prithvi outperforms the other models in segmenting images in unseen regions.”

May combine two claims.

The prediction accuracy (mAcc) of Segformer-B5 is slightly higher than that of Prithvi, and the two models have no obvious performance advantage over the Segformer-B0 model.

“The prediction accuracy (mAcc) of Segformer-B5 is slightly higher than that of Prithvi (Table 2), and the two models have no obvious performance advantage over the Segformer-B0 model.”

• AssumptionRests on an assumption: The 10 runs of each model are sufficient to reliably estimate performance metrics.
• AssumptionRests on an assumption: The six Sentinel-2 bands used as input are sufficient for flood inundation mapping across the studied regions.

May combine two claims.

On the unseen Bolivia data, Segformer-B5 is the second best model, with a difference of less than 1% compared to Prithvi.

“The second best model is Segformer-B5, with a difference of less than 1% compared to Prithvi.”

May combine two claims.

Results show good transferability of the Prithvi model, with performance advantages in segmenting flooded areas in previously unseen regions.

“Results show the good transferability of the Prithvi model, highlighting its performance advantages in segmenting flooded areas in previously unseen regions.”

• AssumptionRests on an assumption: The unseen dataset from Bolivia is representative of previously unseen regions for evaluating the transferability of flood inundation mapping models.
• AssumptionRests on an assumption: The 10 runs of each model are sufficient to reliably estimate performance metrics.
• AssumptionRests on an assumption: The six Sentinel-2 bands used as input are sufficient for flood inundation mapping across the studied regions.

May combine two claims.

The performance of Prithvi and the Segformer models is quite similar in terms of IoU, with a 1% difference in mIoU and nearly a 3% difference in IoU when segmenting flood pixels, compared to U-Net.

“The performance of Prithvi and the Segformer models is quite similar in terms of IoU, with a 1% difference in mIoU and nearly a 3% difference in IoU when segmenting flood pixels, compared to U-Net.”

• AssumptionRests on an assumption: The 10 runs of each model are sufficient to reliably estimate performance metrics.
• AssumptionRests on an assumption: The six Sentinel-2 bands used as input are sufficient for flood inundation mapping across the studied regions.

May combine two claims.

On the unseen Bolivia data, Segformer-B0 ranks third and U-Net ranks lowest, with U-Net nearly 4% lower than Prithvi in both mIoU and mAcc.

“Segformer-B0 follows, while U-Net ranks lowest in terms of performance, with a nearly 4% gap in both mIoU and mAcc compared to Prithvi.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Segformer yields fast processing speed while achieving competitive segmentation results on multiple benchmark datasets.

“Its efficient architecture also yields fast processing speed while achieving competitive segmentation results on multiple benchmark datasets [34].”

• cites [34] 34. Hans Thisanke, Chamli Deshan, Kavindu Chamith, Sachith Seneviratne, Rajith Vidanaarachchi, and Damayanthi Herath. 2023. Semantic segmentation using Vision Transformers: A survey. Engineering Applications of Artificial Intelligence 126 (2023), 106669. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Prithvi's pre-training is based on a Masked AutoEncoder (MAE) strategy.

“The model pre-training is based on a Masked AutoEncoder (MAE) [12] strategy, which asks the model to learn and predict a masked patch from the original image through self-supervised learning.”

• cites [12] 12. Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. 2022. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 16000-16009. sourceNot checked — no full text of the source

May combine two claims.

Sen1Floods11 contains 446 pairs of Sentinel-1 and Sentinel-2 satellite imagery, capturing 11 flood events from 11 countries between 2016 and 2019, with human expert labeled data.

“This dataset contains 446 pairs of Sentinel-1 and Sentinel-2 satellite imagery, which captured 11 flood events from 11 countries (Bolivia, Ghana, India, Nigeria, Pakistan, Paraguay, Somalia, Spain, Sri Lanka, United States, and Vietnam) between 2016 and 2019, with corresponding human expert labeled data.”

• cites [5] 5. Derrick Bonafilia, Beth Tellman, Tyler Anderson, and Erica Issenberg. 2020. Sen1Floods11: A georeferenced dataset to train and test deep learning flood algorithms for sentinel-1. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 210-211. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The final result is presented as the average of each metric obtained from the 10 experiments.

“The final result is presented as the average of each metric obtained from the 10 experiments.”

Classified as probably not a substantial claim. May combine two claims.

U-Net's best performance is attributed to its unique encoder-decoder architecture and its ability to consider multi-scale features to reconstruct spatial details for high-precision segmentation.

“This is attributed to U-Net's unique encoder-decoder architecture and its ability to consider multi-scale features to reconstruct spatial details for high-precision segmentation.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The findings indicate areas for improvement for the Prithvi model in terms of adopting multi-scale representation learning, developing more end-to-end pipelines for high-level image analysis tasks, and offering more flexibility in terms of input data bands.

“The findings also indicate areas for improvement for the Prithvi model in terms of adopting multi-scale representation learning, developing more end-to-end pipelines for high-level image analysis tasks, and offering more flexibility in terms of input data bands.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Prithvi was trained using self-supervised learning on Harmonized Landsat and Sentinel 2 (HLS) data within the continental United States.

“This model was trained using selfsupervised learning on Harmonized Landsat and Sentinel 2 (HLS) data within the continental United States.”

• cites [16] 16. Johannes Jakubik, Linsong Chu, Paolo Fraccaro, Carlos Gomes, Gabby Nyir- jesy, Ranjini Bangalore, Devyani Lambhate, Kamal Das, Dario Oliveira Borges, Daiki Kimura, Naomi Simumba, Daniela Szwarcman, Michal Muszynski, Kommy Weldemariam, Bianca Zadrozny, Raghu Ganti, Carlos Costa, Campbell Edwards, Blair & Watson, Karthik Mukkavilli, Hendrik Schmude, Johannes & Hamann, Parkin Robert, Sujit Roy, Christopher Phillips, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Wei Ji Leong, Ryan Avery, Rahul Ramachan- dran, Manil Maskey, Pontus Olofossen, Elizabeth Fancher, Tsengdar Lee, Kevin Mur sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

U-Net has a simpler architecture than Segformer and Prithvi, so using a slightly larger learning rate helps the model to converge faster.

“As U-Net has a simpler architecture than Segformer and Prithvi, using a slightly larger learning rate helps the model to converge faster.”

Classified as probably not a substantial claim.

For both test and unseen datasets, 10 runs of each model were conducted and performance metrics recorded for each run.

“For both datasets, we conducted 10 runs of each model and recorded the performance metrics for each run.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SAM cannot be easily adopted for semantic segmentation due to differences in the goals of the tasks and the training data used.

“Although SAM provides fine-grained, instance-level segmentation, it cannot be easily adopted for semantic segmentation due to the differences in the goals of these tasks and the training data used.”

• cites [18] 18. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. 2023. Segment anything. arXiv preprint arXiv:2304.02643 (2023). sourceNot checked — no full text of the source

May combine two claims.

Prithvi's underperformance on test data is because U-Net and Segformer architectures enable multi-level feature extraction, whereas Prithvi only utilizes a single-level ViT-based architecture as the feature extraction backbone.

“This is because the U-Net and Segformer architectures enable multi-level feature extraction, whereas Prithvi only utilizes a single-level ViT-based architecture as the feature extraction backbone.”

• Also stated in: § 5 CONCLUSION · plain

  “as it currently only uses single-level features for image analysis, whereas the classic U-Net and the new Segformer model both adopt multi-scale features in their model learning process.”

May combine two claims.

Prithvi's requirement of three additional bands beyond RGB may pose challenges for leveraging image data that do not contain the required bands and therefore negatively affect the model's predictive performance.

“This may pose challenges for leveraging image data that do not contain the required bands and therefore negatively affect the model's predictive performance.”

• AssumptionRests on an assumption: Prithvi requires three additional bands beyond RGB as model input.

May combine two claims.

The geospatial foundation model Prithvi, pre-trained on remote sensing imagery, captures more domain-specific geospatial knowledge, making it particularly suitable for geospatial tasks.

“The geospatial foundation model Prithvi, pre-trained on remote sensing imagery, captures more domain-specific geospatial knowledge, making it particularly suitable for geospatial tasks.”

May combine two claims.

Large models such as Prithvi and Segformer-B5 have a greater capability to learn important feature representations, thus possessing stronger domain adaptability and transferability.

“This result indicates that large models, such as Prithvi and Segfomer-B5, have a greater capability to learn important feature representations, thus possessing stronger domain adaptability and transferability.”

Classified as probably not a substantial claim.

The lack of an end-to-end architecture in Prithvi may limit its adoption by scientists without essential AI knowledge and skills.

“This may limit its adoption by scientists without essential AI knowledge and skills.”

• AssumptionRests on an assumption: The Prithvi model does not include an end-to-end architecture for high-level image analysis tasks, such as object detection or semantic segmentation.