I am currently using the GeoTessera dataset to classify evergreen and deciduous forests in the Amazon basin. However, during my downstream workflow, I encountered noticeable geometric artifacts in the prediction results.
In the attached prediction map, the green pixels represent predicted evergreen trees, and the yellow pixels represent deciduous trees. As you can see, distinct geometric/tiling artifacts propagate through the classification, which noticeably affects the spatial consistency of the results.
I have read in your documentation that slanted artifacts are inherent to Sentinel-1/2 satellite trajectories and can be difficult to eliminate entirely. Given this context, I would like to consult with you on the following questions:
Mitigation/Removal: Is there any recommended preprocessing step, data filtering, or pipeline adjustment on the user end to suppress or eliminate these geometric artifacts?
Impact Minimization: If these artifacts cannot be fully removed, what are the best practices to minimize their negative impact on downstream classification tasks (e.g., feature smoothing, post-processing, or specific model training strategies)?
Any advice or technical guidance you could provide would be greatly appreciated. Thank you very much for your time and support!
I am currently using the GeoTessera dataset to classify evergreen and deciduous forests in the Amazon basin. However, during my downstream workflow, I encountered noticeable geometric artifacts in the prediction results.
In the attached prediction map, the green pixels represent predicted evergreen trees, and the yellow pixels represent deciduous trees. As you can see, distinct geometric/tiling artifacts propagate through the classification, which noticeably affects the spatial consistency of the results.
I have read in your documentation that slanted artifacts are inherent to Sentinel-1/2 satellite trajectories and can be difficult to eliminate entirely. Given this context, I would like to consult with you on the following questions:
Mitigation/Removal: Is there any recommended preprocessing step, data filtering, or pipeline adjustment on the user end to suppress or eliminate these geometric artifacts?
Impact Minimization: If these artifacts cannot be fully removed, what are the best practices to minimize their negative impact on downstream classification tasks (e.g., feature smoothing, post-processing, or specific model training strategies)?
Any advice or technical guidance you could provide would be greatly appreciated. Thank you very much for your time and support!