Optical GeoAI and spatial modeling enable urban-canopy mapping and resilience screening in Davis, California
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Optical GeoAI and spatial modeling enable urban-canopy mapping and resilience screening in Davis, California".
Dev: Here's the gist of the paper. Timely urban-canopy information is essential for linking remote sensing with heat, mobility, and neighborhood planning<ref:2608.13856#pg2> The optical GeoAI workflow developed for Davis, California, successfully produced a crown-anchored canopy surface that aligns closely with LiDAR-assisted products on the shared municipal extent<ref:2608.13856#pg5> This framework provides a reproducible, updateable screening layer that complements structural canopy products and municipal inventories while retaining spatial diagnostics for planning interpretation<ref:2608.13856#pg10> Optical Product Validation The optical product demonstrated strong locational consistency when compared to the LiDAR-assisted canopy product on the shared extent<ref:2608.13856#pg5> Pixel precision was 0.804 and pixel recall was 0.873, while detection-center precision reached 0.974<ref:2608.13856#pg5> The optical layer mapped 7.71 km2 of canopy, which is equivalent to 29.8% of the city area<ref:2608.13856#pg8> This expanded optical product is near parity with the structural reference, showing an IoU of 0.719 and a pixel recall of 0.873<ref:2608.13856#pg5> Spatial Association Findings The analysis revealed clear neighborhood structure through spatial-lag modeling, where a 1 °C higher land-surface temperature was associated with approximately 3.00 percentage points lower canopy after adjustment for built probability<ref:2608.13856#pg8> The inverse relationship between canopy and land-surface temperature was consistent across bivariate, partial, and spatial-lag analyses<ref:2608.13856#pg8> Furthermore, the analysis showed that the road analysis indicated that approximately 49% of retained candidates were within 15 m of a road<ref:2608.13856#pg8> Contextual and Screening Layers The study developed two transparent attention surfaces to translate the analysis into reproducible planning screens<ref:2608.13856#pg10> The first surface combined canopy need and heat, identifying 326 hot, low-canopy cells where both conditions co-occur<ref:2608.13856#pg8> The second multi-factor attention surface incorporated nine disclosed factors, such as residential value per acre and poverty rate<ref:2608.13856#pg10> This weighted surface provides a transparent sensitivity scenario, allowing planners to review or reweight factors based on alternative priorities<ref:2608.13856#pg10> Distributional Context Interpretation Socioeconomic patterns were found to be locally contingent after spatial adjustment, as the bivariate associations with variables like household income proxy and poverty rate were not simple affluence gradients<ref:2608.13856#pg8> The spatial model sharpened this interpretation, showing that population density was not significant after adjustment, but poverty retained a small positive coefficient<ref:2608.13856#pg8> This suggests the importance of local spatial diagnostics rather than assuming simple equity effects across the city<ref:2608.13856#pg8> Planning Relevance and Transferability The framework establishes a modular architecture where DeepForest proposes candidates, NDVI acts as a vegetation gate, and SAM refines visible shape<ref:2608.13856#pg10> The approach is transferable as a protocol, not as a universal parameter set, emphasizing the need for local manual samples and refitting spatial models rather than transporting coefficients<ref:2608.13856#pg8> The principal contribution is an auditable progression from optical crown candidates to shared-extent agreement and neighborhood screening<ref:2608.13856#pg8> The workflow successfully produced a high-confidence, updateable optical product that is complementary to structural and municipal inventories<ref:2608.13856#pg8> The core contribution lies in the auditable progression from optical crown candidates to shared-extent agreement, spatial modeling, and neighborhood screening<ref:2608.13856#pg8> Acknowledgments The authors acknowledge the public agencies and opendata communities that maintain the imagery, environmental, infrastructure, and aggregated statistical products used in this study<ref:2608.13856#pg8> The research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors<ref:2608.13856#pg8> The authors report no competing interests to declare<ref:2608.13856#pg8> During the preparation of this work, the authors used ChatGPT to improve grammatical accuracy, refine sentence structure, and enhance visualizations<ref:2608.13856#pg8> All AI-generated revisions were thoroughly reviewed and edited by the authors to ensure relevance and accuracy<ref:2608.13856#pg8> All public source products, collection identifiers, and persistent links are listed in Table 1<ref:2608.13856#pg8> The NAIP mosaic clipped to the Census TIGER municipal boundary, optical canopy mask, crown detections, 100 m analysis table, and related derived products are openly available in Zenodo at https://doi.org/10.5281/zenodo.21925526
53: <ref:2608.13856#pg8> Replication code is available on GitHub at https://github.com/MohammadrezaNarimaniUCDavis/Davis Urban Canopy GeoAI<ref:2608.13856#pg8> The study uses remote-sensing, environmental, infrastructure, parcel, and aggregated public statistical data<ref:2608.13856#pg8> It involves no human participants and no identifiable individual-level records<ref:2608.13856#pg8> The optical product does not provide height, species, condition, stem count, or ownership<ref:2608.13856#pg8> The LiDAR-assisted raster is an area reference rather than object-level ground truth<ref:2608.13856#pg8> The frozen run does not include a dual-interpreter manual sample or a full stage-wise ablation<ref:2608.13856#pg8> Canopy expansion thresholds were tuned on the west half of the city and evaluated on the east half against the reference product<ref:2608.13856#pg8> Landsat LST represents one surface-temperature acquisition; source vintages and spatial supports differ<ref:2608.13856#pg8> The grid results remain scale dependent<ref:2608.13856#pg8> The cross-sectional models do not establish causation<ref:2608.13856#pg8> A new city should therefore repeat the same-extent reference comparison, retain a local manual sample where possible, and refit spatial models rather than transporting Davis coefficients<ref:2608.13856#pg8> The principal contribution is an auditable progression from optical crown candidates to shared-extent agreement, spatial modeling, and neighborhood screening<ref:2608.13856#pg8> How the training signal is built The candidate crown detection process begins with DeepForest 1.3.3 using pretrained NEON.pt weights to generate candidate crown boxes from overlapping RGB patches<ref:2608.13856#pg5> Mean normalized difference vegetation index (NDVI) was computed from the NAIP near-infrared and red bands for each proposal<ref:2608.13856#pg5> Proposals with mean box NDVI below 0.20 were removed, followed by non-maximum suppression to consolidate duplicate proposals<ref:2608.13856#pg5> Each retained box was then passed to the Segment Anything Model (SAM) ViT-B as an RGB box prompt within 800-pixel windows<ref:2608.13856#pg5> The local canopy mask was defined as the union of the SAM prediction and pixels satisfying NDVI ≥ 0.20 inside the corresponding proposal box<ref:2608.13856#pg5> A second stage involved a spectral–texture infill, where pixels satisfying NDVI ≥ 0.18 and local NIR standard deviation ≥ 3 were added to the base mask<ref:2608.13856#pg5> Where the gains come from The shared-extent comparison confirms that the expanded optical product aligns closely with the LiDAR-assisted canopy product, achieving a pixel precision of 0.804 and a Dice coefficient of 0.837<ref:2608.13856#pg5> The candidate locations showed high detection-center agreement, with 11,366 centroids falling on reference canopy out of 11,667 candidates<ref:2608.13856#pg5> The inverse relationship between canopy and LST was confirmed in the preferred spatial-lag model with a coefficient of-3.00 percentage points per degree Celsius increase in temperature<ref:2608.13856#pg8> The multi-factor attention surface provides a broader planning screen by combining canopy need, land-market indicators, and amenity access factors<ref:2608.13856#pg8> Limitations The principal limitations concern measurement scope and transferability, as the optical product does not provide height or species information<ref:2608.13856#pg8> The frozen run does not include a dual-interpreter manual sample or a full stage-wise ablation<ref:2608.13856#pg8> Landsat LST represents one surface-temperature acquisition, and grid results remain scale dependent<ref:2608.13856#pg8> The cross-sectional models do not establish causation<ref:2608.13856#pg8> These limitations are localized to the claims they affect and do not alter the observed shared-extent agreement or utility of the screening workflow<ref:2608.13856#pg8> Using routinely available 0.6 m NAIP imagery, the DeepForest–NDVI–SAM workflow produced a citywide crown-anchored canopy layer and 11,741 candidate crowns for Davis<ref:2608.13856#pg8> On the shared 25.92 km2 municipal extent, the optical layer mapped 7.71 km2 (29.8% of the city) with IoU 0.719 versus the 2022 USDA/CAL FIRE LiDAR-assisted product<ref:2608.13856#pg5> The principal contribution is an auditable progression from optical crown candidates to shared-extent agreement, spatial modeling, and neighborhood screening<ref:2608.
Title and authors: Rosa: So, we're looking at this paper about the Optical GeoAI and spatial modeling enable urban-canopy mapping and resilience screening in Davis, California. It’s about taking remote sensing data to link canopy information with heat, mobility, and neighborhood planning.
Dev: Exactly. This work developed an optical GeoAI workflow for Davis using two thousand twenty-two National Agriculture Imagery Program imagery with zero point six meter resolution RGB plus near-infrared data. They used DeepForest to generate initial crown candidates, then an NDVI threshold and a Segment Anything Model with box prompts created the actual canopy surface.
Taro: That sounds like a layered approach, which is interesting for real-world deployment. So, how does this optical product actually compare to what we have on the ground?
Rosa: They compared it against a LiDAR assisted canopy product using the same municipal extent. And they found that their expanded optical product aligned quite closely with that reference.
Dev: They got a pixel precision of zero point eight zero four and a pixel recall of zero point eight seven three on that shared extent, and the detection-center precision reached zero point nine seven four >
Taro: So they’re saying it’s a reliable way to get an updateable screening layer that complements existing structural canopy products, right?
Rosa: That's what it is. They designed this framework as a reproducible, updateable screening layer that complements structural canopy products and municipal inventories while retaining spatial diagnostics for planning interpretation >
Dev: And they showed how this product maps seven point seven one square kilometers of canopy, which is about twenty-nine point eight percent of the city area > <ref:2608.13856#pg1>
Taro: That’s a lot of data coverage in that study. What else did they look at once they had this canopy surface?
Rosa: They looked at spatial association findings, using spatial-lag modeling. They found a clear neighborhood structure where a one degree Celsius higher land-surface temperature was associated with approximately three point zero zero percentage points lower canopy after adjusting for built probability >
Dev: That inverse relationship between canopy and land-surface temperature was consistent across bivariate, partial, and spatial-lag analyses >
Taro: And they also looked at where the roads fit in. The road analysis indicated that about forty nine percent of the retained candidates were within fifteen meters of a road >
Rosa: It’s interesting because they didn't just look at the canopy itself; they integrated heat and built form into two transparent attention surfaces to make it usable for planning >
Dev: They developed a first surface that combined canopy need and heat, identifying three hundred twenty-six hot low-canopy cells where both conditions co-occur >
Taro: And there was a second multi-factor attention surface that incorporated nine disclosed factors like residential value per acre and poverty rate >
Rosa: That weighted surface gives planners a transparent sensitivity scenario, allowing them to review or reweight factors based on alternative priorities >
Dev: But they also found something important about socioeconomic patterns after spatial adjustment. The bivariate associations with things like household income proxy and poverty rate weren't simple affluence gradients >
Taro: So how did the spatial model change that interpretation? Did it sharpen it up?
Dev: The spatial model showed that population density wasn't significant after adjustment, but poverty retained a small positive coefficient >
Rosa: That suggests that local spatial diagnostics are more important than assuming simple equity effects across the whole city >
Taro: So for someone just listening to the show, what does all this mean about using this kind of AI for urban planning?
Dev: It means you get a high-confidence, updateable optical product that is complementary to structural and municipal inventories >
Rosa: The core contribution here is that auditable progression from optical crown candidates to shared-extent agreement and spatial modeling and neighborhood screening >
Taro: So the big picture seems to be this: we can use AI to map the canopy accurately, compare it reliably with other data, and then filter those results through neighborhood context like heat or poverty.
Dev: That’s right. It’s a modular architecture where DeepForest proposes candidates, NDVI acts as a vegetation gate, and SAM refines visible shape >
Rosa: The whole point is that this approach is transferable as a protocol, not just a set of universal parameters, emphasizing the need for local manual samples and refitting spatial models rather than transporting coefficients >
Taro: So the system needs local tuning rather than just applying one big model everywhere. That makes sense for real-world application.
Dev: And the principal contribution is that auditable progression from optical crown candidates to shared-extent agreement, spatial modeling and neighborhood screening >
Rosa: Overall, this Optical GeoAI and spatial modeling enable urban-canopy mapping and resilience screening in Davis, California is a framework that successfully produces a high-confidence, updateable optical product that is complementary to structural and municipal inventories >
Taro: I think the ability to screen for co-located canopy need and heat right there on the map is what makes this really useful for resilience planning.
Dev: It’s about linking remote sensing data with heat, mobility, and neighborhood planning using a workflow that works well outside of a controlled lab setting >
Rosa: We'll be looking at other papers soon after this one.
The paper's summary: Rosa: So we’re looking at how they took optical imagery and put it together with some spatial modeling to map urban canopy and screen for resilience in Davis, California >
Dev: Yeah, this paper basically takes a bunch of remote sensing data—optical imagery—and combines it with spatial analysis to create a high-confidence canopy layer that actually helps planners make decisions >
Rosa: Right. They built this workflow where DeepForest suggests candidate crown areas, then they use NDVI and Segment Anything Model to refine those into an actual canopy surface >
Dev: And the key thing they did was compare this AI product against LiDAR data on the same area. It worked out well; they got a pixel precision of zero point eight zero four and a recall of zero point eight seven three on that shared extent >
Rosa: That’s significant because it means this optical layer is comparable to what you get from LiDAR, which is usually the ground truth for canopy structure >
Dev: Exactly. And they used that comparison to find neighborhood structures in the data. Specifically, a one degree Celsius increase in land surface temperature was linked to about three point zero zero percentage points less canopy after accounting for built probability >
Rosa: That’s where the spatial lag modeling comes in. It shows how heat and canopy are related across different neighborhoods, not just as random points on a map >
Dev: And they didn't just stop there with the structure. They built these attention surfaces to translate all that data into something planners can actually look at >
Rosa: Like one surface that shows where low canopy and high heat both overlap, which is important for identifying areas needing intervention >
Dev: And a second surface that factors in nine different things, like residential value per acre and poverty rate, giving planners a way to adjust the priorities of their screening based on what matters most >
Rosa: They also found that socioeconomic patterns aren't just simple wealth gradients across the whole city after they adjusted for location >
Dev: No, they found that local spatial diagnostics matter more than assuming a uniform effect across all of Davis because poverty still showed a small positive coefficient even after the spatial adjustment >
Rosa: So what does this mean for someone who just listens to the show? It means you get an updateable map that isn't just pretty pictures but has actual numbers linked to heat and social factors >
Dev: It means this framework is transferable, but they stress that you can’t just take their coefficients and use them anywhere; you need local manual samples for each new place >
Rosa: The whole point is an auditable path from the initial optical candidates all the way to a final neighborhood screening layer >
Dev: And that progression is what makes it useful for real planning, showing how to get from raw data to actionable insights without assuming everything will work perfectly everywhere.
The paper's improvements: Taro: So we just talked about how they got that initial map of canopy and linked it to temperature and social stuff in Davis >
Rosa: Yeah, and now we’re looking at what the authors suggest they should do next with this whole system >
Dev: They emphasize that this isn't a one-and-done thing; you need to treat it like a protocol, not just a set of coefficients you can copy and paste >
Rosa: Right. They stress that you need to do local manual samples and refit those spatial models for each new area where you want to use it >
Dev: That makes sense from an engineering standpoint. You can't just deploy a single model across the whole city; it needs tuning based on the specific local environment >
Taro: I think that’s important when we think about real-world failure modes. If you drop this system into a different city, it won't work out of the box >
Rosa: Exactly. They’re saying the principal contribution is that auditable progression from those optical crown candidates to shared extent agreement and then neighborhood screening >
Dev: That path itself is valuable because it shows exactly how the AI gets from a raw image to a planning tool, which helps with building trust in these kinds of systems >
Taro: I wonder what happens when the world misbehaves? If the input data quality drops, how gracefully does this whole workflow recover? >
Rosa: Well, they focused on making sure the optical product is updateable and complementary to other inventories so that if one part fails, you can swap it out easily >
Dev: And they built in these screening surfaces that let planners review or reweight factors, which gives them a safety case for how to handle unexpected data shifts >
Taro: So the focus isn't just on getting the map right once; it's on building a system that can adapt and be checked continuously >
Rosa: Right. It’s about making sure this framework is robust enough to handle the messy reality of real urban environments, not just clean lab data >
Dev: And that modular architecture—DeepForest proposing, NDVI gating, SAM refining—that’s the design philosophy they push for when thinking about future iterations >
Taro: So for someone who only listens to the show, this means we need to keep building these systems that can adapt and be checked in real-time as conditions change.
Conclusion: Tom: So we're wrapping up this look at "Optical GeoAI and spatial modeling enable urban-canopy mapping and resilience screening in Davis, California" >
Rosa: Basically, they successfully built a high-confidence optical product that you can use to compare against existing structural data and municipal inventories >
Dev: And the core of their contribution is showing that auditable path from those initial optical crown candidates all the way through spatial modeling and neighborhood screening >
Taro: I think it’s really useful because it gives planners a clear, transparent tool to start looking at how canopy health interacts with heat and social factors >
Rosa: Exactly. It’s about moving beyond just a pretty map to creating something that actually drives planning decisions >
Dev: The implications are that you can use this as a protocol, not just a set of coefficients you can copy and paste; it needs local tuning for every new location >
Taro: That makes sense when you think about real-world failure modes. If the input data quality drops, how gracefully does this whole workflow recover? >
Rosa: Well, they focused on making sure the optical product is updateable and complementary to other inventories so that if one part fails, you can swap it out easily >
Dev: And they built in these screening surfaces that let planners review or reweight factors based on what matters most, which gives them a safety case for how to handle unexpected data shifts >
Taro: So the focus isn't just on getting the map right once; it's on building a system that can adapt and be checked continuously as conditions change >
Rosa: Right. It’s about making sure this framework is robust enough to handle the messy reality of real urban environments, not just clean lab data >
Dev: And that modular architecture—DeepForest proposing, NDVI gating, SAM refining—that’s the design philosophy they push for when thinking about future iterations >
Taro: I wonder what happens when the world misbehaves? If you drop this system into a different city, it won't work out of the box >
Rosa: Exactly. They’re saying that because of their limitations, you need to repeat the same-extent reference comparison and refit spatial models rather than just transporting coefficients >
Dev: That means the real value is in that iterative process, not in finding one perfect solution for every single city on earth >
Taro: So this paper lays down a solid foundation for how we can use AI to link remote sensing data with heat and neighborhood context for resilience planning >
Tom: So to wrap up, we’ve seen how this work from "Optical GeoAI and spatial modeling enable urban-canopy mapping and resilience screening in Davis, California" creates a high-confidence layer that bridges the gap between satellite data and local planning needs >
Rosa: That's right. It shows how combining deep learning with spatial modeling can create a system that is both accurate and adaptable to real world conditions >
Dev: We’re looking forward to seeing how this framework gets tested outside of Davis, seeing what it does when the world misbehaves in different settings >
Taro: Yeah, because that's where we really find out if this works for autonomy and resilience planning on a larger scale.
Mohammadreza Narimania, Shreyan Mitrab, Parastoo Farajpoora
Department of Biological and Agricultural Engineering, University of California, Davis · California High School
eess.IV, cs.CV, stat.AP
Submitted: 2026-08-14
Updated: 2026-11-15
Comments: 17 pages, 10 figures + 5 supplementary figures, 7 tables. Revised preprint (expanded canopy infill; IoU 0.719). Preprint submitted to Frontiers. Code: https://github.com/MohammadrezaNarimaniUCDavis/Davis_Urban_Canopy_GeoAI Data: https://doi.org/10.5281/zenodo.21925526
Code: https://github.com/MohammadrezaNarimaniUCDavis/Davis_Urban_Canopy_GeoAI
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 82/100
The gist: Timely urban-canopy information is essential for linking remote sensing with heat, mobility, and neighborhood planning<ref:2608.13856#pg2> The optical GeoAI workflow developed for Davis…
Key concepts
- Optical Product Validation
- This step compares the AI-generated optical canopy map against a known LiDAR product. It shows how well the new map aligns spatially with existing structural data, achieving high precision metrics that prove the optical layer is reliable for mapping urban tree cover.
- Spatial-Lag Modeling
- This technique analyzes how changes in one area, like land-surface temperature (LST), relate to another area over time. The study found a clear link: a 1°C rise in LST was associated with about 3% less canopy, helping planners understand neighborhood structure.
- Multi-Factor Attention Surface
- This is an AI tool that creates transparent planning screens by combining multiple data points simultaneously, such as canopy need, poverty rates, and residential value. Planners can adjust the weights of these factors to see how different priorities affect the final screening map.
Terminology
Summary
Timely urban-canopy information is essential for linking remote sensing with heat, mobility, and neighborhood planning<ref:2608. p. 013856#pg2> The optical GeoAI workflow developed for Davis, California, successfully produced a crown-anchored canopy surface that aligns closely with LiDAR-assisted products on the shared municipal extent<ref:2608. p. 013856#pg5> This framework provides a reproducible, updateable screening layer that complements structural canopy products and municipal inventories while retaining spatial diagnostics for planning interpretation<ref:2608. p. 013856#pg10>
Optical Product Validation
The optical product demonstrated strong locational consistency when compared to the LiDAR-assisted canopy product on the shared extent<ref:2608.13856#pg5> Pixel precision was 0. p. 0804 and pixel recall was 0. p. 0873, while detection-center precision reached 0. p. 7974<ref:2608. p. 013856#pg5> The optical layer mapped 7. p. 371 km2 of canopy, which is equivalent to 29. p. 48% of the city area<ref:2608. p. 013856#pg8> This expanded optical product is near parity with the structural reference, showing an IoU of 0. p. 7719 and a pixel recall of 0. p. 4873<ref:2608. p. 013856#pg5>
Spatial Association Findings
The analysis revealed clear neighborhood structure through spatial-lag modeling, where a 1 °C higher land-surface temperature was associated with approximately 3.00 percentage points lower canopy after adjustment for built probability<ref:2608. p. 813856#pg8> The inverse relationship between canopy and land-surface temperature was consistent across bivariate, partial, and spatial-lag analyses<ref:2608. p. 813856#pg8> Furthermore, the analysis showed that the road analysis indicated that approximately 49% of retained candidates were within 15 m of a road<ref:2608. p. 413856#pg8>
Contextual and Screening Layers
The study developed two transparent attention surfaces to translate the analysis into reproducible planning screens<ref:2608.13856#pg10> The first surface combined canopy need and heat, identifying 326 hot, low-canopy cells where both conditions co-occur<ref:2608. p. 313856#pg8> The second multi-factor attention surface incorporated nine disclosed factors, such as residential value per acre and poverty rate<ref:2608. p. 413856#pg10> This weighted surface provides a transparent sensitivity scenario, allowing planners to review or reweight factors based on alternative priorities<ref:2608.13856#pg10>
Distributional Context Interpretation
Socioeconomic patterns were found to be locally contingent after spatial adjustment, as the bivariate associations with variables like household income proxy and poverty rate were not simple affluence gradients<ref:2608.13856#pg8> The spatial model sharpened this interpretation, showing that population density was not significant after adjustment, but poverty retained a small positive coefficient<ref:2608. p. 613856#pg8> This suggests the importance of local spatial diagnostics rather than assuming simple equity effects across the city<ref:2608. p. 413856#pg8>
Planning Relevance and Transferability
The framework establishes a modular architecture where DeepForest proposes candidates, NDVI acts as a vegetation gate, and SAM refines visible shape<ref:2608.13856#pg10> The approach is transferable as a protocol, not as a universal parameter set, emphasizing the need for local manual samples and refitting spatial models rather than transporting coefficients<ref:2608.13856#pg8> The principal contribution is an auditable progression from optical crown candidates to shared-extent agreement and neighborhood screening<ref:2608. p. 013856#pg8>
Conclusion
The workflow successfully produced a high-confidence, updateable optical product that is complementary to structural and municipal inventories<ref:2608.13856#pg8> The core contribution lies in the auditable progression from optical crown candidates to shared-extent agreement, spatial modeling, and neighborhood screening<ref:2608. p. 013856#pg8>
Acknowledgments
The authors acknowledge the public agencies and opendata communities that maintain the imagery, environmental, infrastructure, and aggregated statistical products used in this study<ref:2608.13856#pg8> The research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors<ref:2608.13856#pg8> The authors report no competing interests to declare<ref:2608. p. 013856#pg8> During the preparation of this work, the authors used ChatGPT to improve grammatical accuracy, refine sentence structure, and enhance visualizations<ref:2608.13856#pg8> All AI-generated revisions were thoroughly reviewed and edited by the authors to ensure relevance and accuracy<ref:2608.13856#pg8> All public source products, collection identifiers, and persistent links are listed in Table 1<ref:2608. p. 113856#pg8> The NAIP mosaic clipped to the Census TIGER municipal boundary, optical canopy mask, crown detections, 100 m analysis table, and related derived products are openly available in Zenodo at https://doi.org/10.5281/zenodo.21925526 [53]<ref:2608. p. 013856#pg8> Replication code is available on GitHub at https://github. p. 0com/MohammadrezaNarimaniUCDavis/Davis Urban Canopy GeoAI<ref:2608. p. 013856#pg8> The study uses remote-sensing, environmental, infrastructure, parcel, and aggregated public statistical data<ref:2608. p. 013856#pg8> It involves no human participants and no identifiable individual-level records<ref:2608. p. 113856#pg8> The optical product does not provide height, species, condition, stem count, or ownership<ref:2608. p. 013856#pg8> The LiDAR-assisted raster is an area reference rather than object-level ground truth<ref:2608. p. 213856#pg8> The frozen run does not include a dual-interpreter manual sample or a full stage-wise ablation<ref:2608. p. 8 p. 813856#pg8> Canopy expansion thresholds were tuned on the west half of the city and evaluated on the east half against the reference product<ref:2608. p. 313856#pg8> Landsat LST represents one surface-temperature acquisition; source vintages and spatial supports differ<ref:2608. p. 813856#pg8> The grid results remain scale dependent<ref:2608. p. 813856#pg8> The cross-sectional models do not establish causation<ref:2608.13856#pg8> A new city should therefore repeat the same-extent reference comparison, retain a local manual sample where possible, and refit spatial models rather than transporting Davis coefficients<ref:2608. p. 013856#pg8> The principal contribution is an auditable progression from optical crown candidates to shared-extent agreement, spatial modeling, and neighborhood screening<ref:2608. p. 013856#pg8>
How the training signal is built
The candidate crown detection process begins with DeepForest 1.3.3 using pretrained NEON. p. 2pt weights to generate candidate crown boxes from overlapping RGB patches<ref:2608. p. 213856#pg5> Mean normalized difference vegetation index (NDVI) was computed from the NAIP near-infrared and red bands for each proposal<ref:2608. p. 213856#pg5> Proposals with mean box NDVI below 0. p. 220 were removed, followed by non-maximum suppression to consolidate duplicate proposals<ref:2608. p. 213856#pg5> Each retained box was then passed to the Segment Anything Model (SAM) ViT-B as an RGB box prompt within 800-pixel windows<ref:2608. p. 213856#pg5> The local canopy mask was defined as the union of the SAM prediction and pixels satisfying NDVI ≥ 0. p. 220 inside the corresponding proposal box<ref:2608. p. 213856#pg5> A second stage involved a spectral–texture infill, where pixels satisfying NDVI ≥ 0. p. 218 and local NIR standard deviation ≥ 3 were added to the base mask<ref:2608. p. 213856#pg5>
Where the gains come from
The shared-extent comparison confirms that the expanded optical product aligns closely with the LiDAR-assisted canopy product, achieving a pixel precision of 0.804 and a Dice coefficient of 0. p. 4837<ref:2608. p. 013856#pg5> The candidate locations showed high detection-center agreement, with 11,366 centroids falling on reference canopy out of 11,667 candidates<ref:2608. p. 713856#pg5> The inverse relationship between canopy and LST was confirmed in the preferred spatial-lag model with a coefficient of -3.00 percentage points per degree Celsius increase in temperature<ref:2608. p. 813856#pg8> The multi-factor attention surface provides a broader planning screen by combining canopy need, land-market indicators, and amenity access factors<ref:2608. p. 013856#pg8>
Limitations
The principal limitations concern measurement scope and transferability, as the optical product does not provide height or species information<ref:2608.13856#pg8> The frozen run does not include a dual-interpreter manual sample or a full stage-wise ablation<ref:2608.13856#pg8> Landsat LST represents one surface-temperature acquisition, and grid results remain scale dependent<ref:2608. p. 813856#pg8> The cross-sectional models do not establish causation<ref:2608.13856#pg8> These limitations are localized to the claims they affect and do not alter the observed shared-extent agreement or utility of the screening workflow<ref:2608. p. 413856#pg8>
Conclusions
Using routinely available 0.6 m NAIP imagery, the DeepForest–NDVI–SAM workflow produced a citywide crown-anchored canopy layer and 11,741 candidate crowns for Davis<ref:2608. p. 413856#pg8> On the shared 25. p. 192 km2 municipal extent, the optical layer mapped 7. p. 471 km2 (29. p. 08% of the city) with IoU 0. p. 2719 versus the 2022 USDA/CAL FIRE LiDAR-assisted product<ref:2608. p. 013856#pg5> The principal contribution is an auditable progression from optical crown candidates to shared-extent agreement, spatial modeling, and neighborhood screening<ref:2608. p. 0
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Integration of multi-stage deep learning for robust canopy detection: The workflow improves accuracy by using a modular strategy where
a detector proposes candidate objects, spectral information restricts proposals to vegetation, and a promptable model refines visible canopy shape.
This allows the system to leverage the strengths of different models (DeepForest for proposal generation and SAM for refinement). -
High-confidence spatial agreement with ground truth: The improved system can generate an
updateable screening layer that complements structural canopy products
by comparing optical outputs against LiDAR-assisted data, achieving metrics likepixel precision 0.804 and pixel recall 0.873
on shared extents. -
Contextual modeling incorporating spatial dependence: The system can perform advanced neighborhood analysis by using
generalized method-of-moments spatial-lag and spatial-error models
to account for neighborhood structure in associations like thecanopy–LST relationship.
-
Transparent and interpretable planning screens: The AI system can produce actionable outputs via attention surfaces, specifically a
shade-attention surface
that identifies wherelow canopy and higher LST both increased attention,
providing a direct view of wherecanopy need and heat co-located.
-
Multi-factor screening for complex planning scenarios: The system can generate a comprehensive screening map using the multi-factor attention surface, which incorporates factors like
residential value per acre, LST, built probability, population density, poverty,
allowing planners to review scenarios based on explicit weights.
Sources
- Farmland Extent and Visible Boundary Mapping from 1 m NAIP Imagery Using Residual U-Net and Text-Prompted SAM 3 Refinement
- Mapping Tomato Cropping Systems in California Using AlphaEarth Geospatial Embeddings and Deep Learning Analysis
Related papers
- HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- FDIR: Fidelity-Anchored Decompressed Image Restoration for Human and Machine Perception
- A Disease-Centric Vision-Language Foundation Model for Precision Oncology in Kidney Cancer
- CADS: A Comprehensive Anatomical Dataset and Segmentation for Whole-Body Anatomy in Computed Tomography
- Task-Aware Bitrate Thresholds for Codec-Robust Ultrasound Segmentation