伟吉

Giving a talk on 7 Nov 2022 at 17:00 (UTC) about ☯️ #zen3geo: Guiding Earth Observation 🌏 data on its path to enlightenment. Learn about #torchdata #DataPipes for #geospatial #machinelearning! More info at discourse.pangeo.io/t/monday-novem… P.S. I'm now over at @weiji14" target="_blank" rel="nofollow noopener">mastodon.nz/@weiji14

@rybavery @EricFielding @Ak_Satellite @GEDI_Knights I see. If you're doing biomass estimation though, won't you still need to combine L-band (which gives you ground) and X/C/P-band (which gives you canopy)? Assuming the land is stable, maybe requesting a few L-band images from JAXA (or a mosaic) would suffice?

@weiji14_ @EricFielding @Ak_Satellite @GEDI_Knights Thanks good suggestion. we looked at daac.ornl.gov/GEDI/guides/GE… but the resolution isn't sufficient. 1kmx1km when we need 1 hectare resolution. Also GEDI may not be around long term so we'd prefer to use L-band SAR, since there are upcoming L and P-band SAR missions

anybody know what's the easiest way to access #palsar-2 L-band imagery? Best I can find is this, annual mosaics, but it doesn't include individual images for sub-annual work: twitter.com/DEarthAfrica/s…

@rybavery @EricFielding @Ak_Satellite Ah ok, have you also looked at @GEDI_Knights? That would be open data, though it's a point cloud format.

@weiji14_ @EricFielding @Ak_Satellite Thanks for the recs! It'd be for quarterly monitoring of above ground biomass in regrowing forest in the South American tropics. Very much looking forward to NISAR launch, and I'm hoping it'll be more accessible. And BIOMASS as well for monitoring changes in denser forest.

@GreyNearing @kbgg_ @OurRadiantEarth @GeoSkeptic ML-ready data would be a level-up from Analysis Ready Data, think Level 3+, something that is ready to be loaded into a Tensor object.

@GreyNearing @kbgg_ @OurRadiantEarth Agree, @GeoSkeptic wrote a nice blog post about the nuances of Analysis Ready Data at element84.com/blog/deconstru…. The idea of ARD is to have 'some' pre-processing done, but how much or how far is use-case dependent.

@rybavery @alkalait @sat_summit @marth_hm @geonanayi @ImpactObserv @xarray_dev @pangeo_data Oh, and I'm planning to give a short presentation about #zen3geo on 7 Nov 2022 (12 noon, Eastern Time) at the Pangeo ML Working group meeting (#working-group-meetings" target="_blank" rel="nofollow noopener">pangeo.io/meeting-notes.…). Will be fleshing out some ideas and hoping to get a few features in over October.

@rybavery @alkalait @sat_summit @marth_hm @geonanayi @ImpactObserv No problem, my ears actually perked up when you mentioned during the panel how #ML4EO needs to coalesce around the @xarray_dev data model. zen3geo simply glues together standard @pangeo_data tools, and I'm excited to see it enable multi-modal/multi-dimensional applications 😀

Cheers @alkalait for moderating @sat_summit's AI Production - Tricks and Tools session with Carolyn Johnston (DevGlobal Ramp), @marth_hm, @rybavery, @geonanayi & Megan Hansen (@ImpactObserv). Here's a 🎥 screen recording (sry for halfway start and lag) 🙈 youtu.be/paAYHZsdbXM

Only managed to save 3.5 recordings for Day 2 #SatSummit2022. Pls pardon any missing bits, need a more cloud-optimized livestream 🙂 Thanks again to @developmentseed, dev.global, @Cyient and everyone for a great @sat_summit! Full playlist: youtube.com/playlist?list=…

@sat_summit @laurenallognon @ninafikiri @jawoo @IFPRI @AidData @corteva @htkilic @ridwansorunke @weathergaige @earthhacksorg @Mabel_Bae Perspectives from cutting edge applications of #EO data in forecasting & prediction @sat_summit. Carolyn Johnston (DevGlobal Ramp) moderating the session with @BudhuBhaduri, Rahul Ramachandran (@NASA_Marshall), @abitmore (@cloud2street) & @jawoo (@IFPRI). youtu.be/hmzweJh1y0s

Panel session on Feeding the Planet 🌍🌏🌎 @sat_summit. @laurenallognon moderating the session with @ninafikiri, @jawoo (@IFPRI), Kunwar Singh (@AidData), Anu Swatantran (@corteva) and @htkilic. youtu.be/MRVJmUqiyJw

@cavrinceanu @sat_summit I'm screen recording some and will put them up on this playlist - youtube.com/watch?v=paAYHZ…. Only just started halfway this morning on Day 2 though 🙂

@sat_summit Will there be any recording of this livestream? I missed a few interesting sessions. 😣

Your invitation to join us! #SatSummit2022 youtu.be/L3Q8Yfk092U

@alkalait @sat_summit Haha that's ok, the speakers seemed more optical focused, though Ryan did mention some Sentinel-1. Great job moderating by the ways, lots of philosophical questions!

@weiji14_ @sat_summit Damnit. I saw that too late.

Tomorrow @sat_summit I'll host "AI Production: Tricks & Tools" To virtual attendees: what would like to ask our panelists? #29-1000-ai-production-tricks-tools" target="_blank" rel="nofollow noopener">2022.satsummit.io/agenda#29-1000…

@dinomirMT @Glennwithtwons Ah, should have mentioned object detection in the first place! Then look into github.com/obss/sahi which is designed for large images, cc @fcakyon. Once you've got the boxes, georeference them following zen3geo.readthedocs.io/en/v0.5.0/obje…

@weiji14_ @Glennwithtwons I have a super high res (10cm) dataset over 30x30km. Looking for ways to segment and run object detection. Edge artefacts are exactly what I want to avoid. May need some more pointers to model sharding ... thanks!

#ML / #COG Tweeps, seeking pointers to any resources on training/inference on large # spatial #rasters (COGs, on S3). Tiling the data seems a step back. Can we stream in (parallel?) parts at full res, accessing a range? Seems like the right way, any pointers? #pytorch #S3

@dinomirMT @Glennwithtwons Model sharding will help to avoid tiling edge artifacts when running inference (though it works for training too). But yes, tiling/chipping is still ok, especially if your training data is limited to a few satellite scenes in a small geographic regions.

@Glennwithtwons @weiji14_ Not sure I understand model sharding, I think the data tiling (although not optimal) will work. Will explore the sharding too, though.

@Glennwithtwons @dinomirMT On the contrary, I'd recommend looking into model sharding rather than data sharding (of which tiling is one way). This is something the #ML4EO community needs to pick up on more, see e.g. using deepspeed at devblog.pytorchlightning.ai/experiment-wit…

@dinomirMT this might describe what you are looking for: zen3geo.readthedocs.io/en/v0.5.0/chip… @weiji14_

☯️ #zen3geo v0.5.0. Cloud-native #ML4EO is here, say hello to @STACspec @Pytorch #DataPipes! Here's a #Sentinel-1 #SAR time-series 📆 + @CopernicusEU #DEM landslide change-detection 🕵️ walkthrough over Gunung Talamau, Indonesia 🇮🇩. zen3geo.readthedocs.io/en/v0.5.0/stac…. Ah, #GoldenAgeOfSAR 🌅