Michael Dayah

Earth observation operators face a strategic fork. Scale the fleet. Add more spacecraft with different sensors, different orbits, different tasking profiles. More coverage, more revisit options, but you're compounding heterogeneity. Maxar's evolution from WorldView to Legion is the template. Standardize into a constellation. The Planet model: uniform satellites, predictable operations, economies of scale. Massive architectural bet that doesn't suit every mission profile. Connect to orbital compute. This is the emerging third path. Starcloud recently demonstrated training a small LLM in orbit on an H100-class payload. Axiom launched Orbital Data Center nodes to LEO earlier this month, positioning space-based processing that can connect to other spacecraft via compatible links. Keep your existing fleet, process on-orbit, downlink insights not raw data. The strategic question shifts: build, buy, or rent orbital compute capacity? #EarthObservation #SatelliteOperations #NewSpace #OnOrbitCompute

52

The drone industry is obsessed with products. Airframes. Payloads. Specs. “Agile.” “Iterative.” “Game-changing.” But the uncomfortable truth is this: drones don’t win — people do. The battle-winning advantage isn’t the platform. It’s the operator + technician skillset: Mission literacy (knowing the “why,” not just the route) RF/EW awareness (how to work when the spectrum gets ugly) Sensor craft (collecting what’s useful, not just what’s visible) Fault-finding + sustainment (keeping tempo) The ability to adapt TTPs fast We keep buying “solutions” as if the problem will stay still. It won’t. The adversary adapts, the environment changes, and today’s best kit becomes tomorrow’s liability. A skills-first model also creates longevity: when this drone era moves on, you still have highly capable people—technical, tactical, and adaptable—who transfer into comms, ISR, engineering, training, and leadership. Stop asking “Which drone?” Start asking “What kind of drone people are we building?” A trimmed version of a much bigger article.

34

3 Comments

The most valuable drone professionals I know rarely touch a drone. Not because they don’t know how to fly. But flying is the least valuable part of the workflow. The people who actually drive outcomes in this industry spend their time on: GIS and spatial analysis remote sensing interpretation ISR-style workflows data validation and QA translating raw imagery into decisions In other words, they work on meaning, not machinery. Anyone can learn to fly a drone in a weekend. Very few people can answer: What does this data actually represent? What’s the uncertainty? What decisions can safely be made from this? That gap is where real value lives. It’s also why so many drone businesses struggle. They optimize: flight time sensors platforms But neglect: analytics pipelines domain expertise client integration operational context The market doesn’t pay for airtime. It pays for insight. Drones are just the newest sensor in a much older discipline: remote sensing + GIS + systems thinking. The professionals who understand that stop being “drone pilots.” And start becoming a decision infrastructure for real industries.

168

48 Comments

🚴‍♂️ Joining 53 million CitiBike trips to NYC neighborhoods in Apache Sedona on my laptop. Yesterday I used Apache Sedona to run a spatial join between tens of millions of Citibike trip origins and NYC neighborhood boundaries and it handled it beautifully. If you’ve ever tried doing this in GeoPandas, you know memory limits hit fast. But thanks to the work of some amazing colleagues and an awesome blog post from Dewey Dunnington and a tutorial by Matthew Powers, CFA, which helped me get Spark + Sedona set up super fast. Here’s how that worked: 🔧 Set up with Sedona was smoother than expected using SDKMAN to manage Java, although the Docker image install of Sedona works as well. 📦 Data: I loaded 53M records of Citibike trip start points from Source Cooperative and NYC neighborhood polygons into Sedona dataframes, cast geometries with ST_Point, and used ST_Contains for the spatial join. ⚙️ Speed boost: I added some simple repartitions to spread the data around the mini Spark cluster and time to improve performance. ✅ End result: As you can see on the map this counts every trip that started in each NYC neighborhood - and I turned it into a GeoJSON file by moving the Sedona dataframe to Geopandas with two lines of code. Whether you’re dealing with millions of features or just tired of MemoryError messages, Sedona is a great tool to add to your spatial stack. 🌎 I'm Matt and I talk about modern GIS, geospatial data engineering, and how spatial thinking is changing. 📬 Want more like this? Join 5k+ others learning from my newsletter → forrest.nyc

128

22 Comments

🤖 JAX port of the famous PointMaze environment from Gymnasium Robotics! I built this for my own research and thought it might also be helpful to fellow researchers. Nothing groundbreaking, but the JAX implementation delivers millions of environment steps per minute with full JIT/vmap support. Perfect for anyone doing navigation research, goal-conditioned RL, or just needing fast 2D maze environments. Plus, easy custom maze creation from simple 2D layouts! Feel free to contribute and drop a star ⭐️! 🔗 https://lnkd.in/euPNDj8D #JAX #ReinforcementLearning

61

1 Comment

Ever tried stitching together street-level property photos and drone shots? Probably not because that's a pretty wild idea. At CVPR 2025, researchers introduced AerialMegaDepth, a new dataset and training method that fuses aerial images with ground-level photos to build detailed 3D reconstructions of real-world scenes. This model bridges the gap between Google Earth-style aerial views and street imagery—and it’s pretty amazing. Earlier models could only align around 5% of aerial-ground views accurately. AerialMegaDepth hit over 56%. That’s not a small improvement—it’s a breakthrough. This kind of leap opens up massive opportunities. Think: richer virtual walkthroughs, further enhanced automated valuations, and more accurate damage assessments from every angle—top-down and curbside. This is the kind of research that makes the future of real estate even more exciting. https://lnkd.in/eAdmmq-B [2504.13157] AerialMegaDepth: Learning Aerial-Ground Reconstruction and View Synthesis #CVPR2025 #ComputerVision #3DReconstruction #PropTech #FoxyAI

9

My favorite thing about working in SSA is that I can use the worst case study and still be right. In 2025, Starlink executed ~300,000 collision avoidance maneuvers. With roughly ~9,000 operational satellites, that’s ~33 maneuvers per satellite per year. Over a 7–8 year theoretical hardware upper bound, that implies ~230–260 avoidance maneuvers per satellite lifetime. Operationally, Starlink satellites tend to be retired closer to ~5 years, even though hardware survival could plausibly extend toward ~7–8 years under favorable conditions. If we use an upper bound and assume up to ~30% of lifetime Δv is allocated to collision avoidance (with the remaining ~70% going to orbit raise, drag make-up, stationkeeping, constellation maintenance, and disposal), then CA represents a meaningful fraction of total mission consumables. Now consider the quality of orbital knowledge. If conjunction assessments are being generated from orbit solutions with ~200 m RMS uncertainty (hypothetical scenario), Pc estimates become materially inflated relative to tighter covariance. At fixed maneuver thresholds (e.g., 1e-6 or lower), this mechanically increases maneuver frequency. Even if only a modest fraction of CA maneuvers were driven by covariance overestimation rather than true high-risk conjunctions, that would mean a non-trivial portion of that ~30% Δv allocation is being consumed on events that might not have required action under tighter uncertainty bounds. BUT the funny part to me is that Starlink hides behind a very conservative maneuvering strategy: operating on a CA for a Pc of 3:1,000,000. Imagine if every satellite listened to the standard operational procedure of 1:10,000. I’m saying all of this to make one very simple claim: We do way too much collision avoidance. And the reason for that lies in the fact that all companies in SSA right now are focused on either tracking smaller debris or doing covariance modeling. Me? Screw all of that. You have to get the first job done to it’s completion before you worry about the next. So yes, xOrbita Inc will be building for the future and staying behind to clean up the mess SSA has already made. (Amanda Fadely , Dr Adam E. White , John Avera , E2MC (Earth-to-Mars Capital) , Sigrid Elschot) #SSA #Space #SpaceDebris #startup #starlink #spacex

69

15 Comments

One missing paragraph in a contract can kill a drone program. We’ve seen it happen more than once. Everything looked ready: certified pilots, compliant aircraft, clear mission goals, and still the program stalled. Not because the technology failed. Because the agreement wasn’t built for operations. 1️⃣ Compliance must be written, not assumed Part 107. Aircraft registration. Remote ID. If it isn’t explicitly required in the agreement, you’re operating on trust, not governance. 2️⃣ Data ownership decides everything later Before the first flight, ask: Who owns the data? Where is it stored? Who can access it? How long is it kept? This is one of the biggest hidden risks in municipal programs. 3️⃣ Deliverables > flight hours “Fly missions” is not an outcome. Define: • File formats • Accuracy standards • GIS/asset system integration Otherwise, you end up with data nobody can use. 4️⃣ Cybersecurity is now an operations issue Drone data often includes sensitive infrastructure. Contracts should cover: • Encryption • Access controls • Breach response • Limits on personal data capture This isn’t optional anymore. 5️⃣ Insurance clarity protects the program Every partner should carry: • Drone-specific liability • Workers’ comp • Public-sector aligned indemnification You don’t want to figure this out after an incident. 6️⃣ Sub-contractors are where compliance breaks The main vendor is compliant. Sub-vendor not. Make sure every requirement flows down the chain. Big lesson: Strong agreements don’t slow drone programs down. They’re what allow them to scale safely, legally, and sustainably. The real work of drone operations starts long before the first flight. It starts on paper.

52

5 Comments

🚁 Drone Pilots in Charlotte – Big News! 🛫 ‼️ New LAANC Grids Are Live! ‼️ Flying around the Charlotte area just got a whole lot easier. The updated grids mean more flexibility, better access, and smoother mission planning for certified drone pilots. What this means for us: Easier route planning near controlled airspace More opportunities for aerial projects Clearer altitude limits and safer operations Whether you’re capturing real estate, inspecting infrastructure, or creating stunning aerial content, these changes give us the space and confidence to operate safely and efficiently. Pro Tip: Check your favorite LAANC ( Low Altitude Authorization and Notification Capability) -enabled app (Aloft, AirMap, UASidekick, etc.) to see the new altitudes and get instant authorizations for your next mission. Charlotte skies are opening up — let’s continue to keep them safe! Fly Responsibly! You can get FAA LAANC maps for your drone mission directly from the FAA’s UAS Facility Maps page. Here’s the official link: https://lnkd.in/evGGwHM9 #LessonsFromSHE #HotDroneJob #AirHustleCollabs #SkyTrapperCollabs #CollabInAction #Part107Collabo #OpenSkyImageries #CLTAirportOperations

36

Exit velocity? The drone delivery industry might finally be getting there. In just the past few weeks we’ve seen milestone after milestone: The FAA’s long-awaited Part 108 proposal is on the table — moving BVLOS from waiver-based exceptions toward a repeatable framework. https://lnkd.in/efgS5yi5 Walmart has already completed 300,000 deliveries, hitting 1,786 in a single day, and is expanding drone delivery to 6 metros and 100 stores this year. https://lnkd.in/ejvy6NYr Amazon just opened a new Prime Air hub in Florida, with more metros planned for later this year. https://lnkd.in/ejiyJGQA Uber announced a partnership with Flytrex. https://lnkd.in/ec86jTkq At the All-In Summit, Uber’s CEO predicted that half of food deliveries will shift to drones or sidewalk bots. https://lnkd.in/ecmmG3Ne Also at the Summit, Zipline’s CEO gave a presentation worth watching — laying out the science-meets-execution story of what scaled operations can really look like. https://lnkd.in/ey46Ga7u So yes, worth celebrating: the industry is moving in the right direction. But let’s not forget the one question almost nobody wants to be very loud about: cost per package. For drone delivery to move from hype to everyday necessity, it has to become economically sustainable. Not a small challenge — but the next one to watch, and the one that will ultimately decide who wins this race.

39

4 Comments

I've been struggling to understand D.C.'s totally revamped bus network so I grabbed the GTFS data and made a simplified little route map. Turns out there is a pattern to the numbers! Analysis + mapping all done in R + a tiny bit of design in Inkscape. And an experiment with Quarto's blog template: mostly nice but needs better options for inserting HTML. https://lnkd.in/ee56qpM7

17

A couple of weeks ago, Nick Horob proposed the idea of building a tool that would analyze the sentiment of the Market Talk forum of AgTalk (for non-ag folks, it's an old school forum that's a bastion of poorly documented knowledge). It just so happens that I've been scraping data from AgTalk for nearly a year now, so I already had the data. https://agtalkpulse.com/ My biggest surprise from building it was how neutral the average sentiments are! Here's a quick breakdown of how it works: 1) A crawler consistently checks the active pages of AgTalk's Market Talk forum, scraping and persisting the posts. 2) When a new post is persisted, the Anthropic structured data API is used to assess a sentiment score and extract crop mentions and category tags. 3) A simple NextJS dashboard exposes the data for exploration. If you have any suggestions or feedback, feel free to share it! This was really just a small toy project so I could use my NAS a bit more! I'm happy to share the data with anyone who is interested.

34

8 Comments