Companion Guide — Hiyagon Mangrove Satellite & LiDAR Report

This document accompanies hiyagon_report.html. It explains, in detail, every figure in the report, the technical methods behind them, and how each result compares with the published literature. Inline citations point to the numbered References at the end.

The analysis maps mangrove extent, change, biomass and carbon over the Hiyagon wetland (Okinawa, Japan; ~26.31 °N, 127.83 °E) by combining four independent data sources, all reprojected to EPSG:32652 (WGS84 / UTM 52 N).

1. Data sources

The AOI is the convex hull of the 13 polygons (~4.4 ha, ~444 pixels at 10 m).

2. Methods (the technical sections, explained)

2.1 Shared-model clustering and change detection

Each annual data stack (AlphaEarth 64-band, or Sentinel-2 15-band) is clustered with k-means [7,8]. One model is fit on all hull pixels pooled across all years (features standardised to zero mean / unit variance), then applied per year, so cluster identities stay consistent between years — a prerequisite for per-pixel change detection. The number of clusters k (3–7) is chosen by the silhouette coefficient [9] (k = 3 here for both stacks). The cluster whose pixels most overlap the 13 prior polygons is labelled “mangrove” (anchoring the unsupervised result to expert ground truth).

Change is computed per pixel by comparing the mean mangrove state of the first three vs last three years: a pixel is growth (non-mangrove → mangrove), decay (mangrove → non-mangrove), stable mangrove, or non-mangrove.

2.2 Sentinel-2 spectral features

Sentinel-2’s finest native resolution is 10 m; the 20 m bands add spectral detail (not finer pixels) and are resampled to 10 m. The 15-band feature stack is the 10 m bands (B2, B3, B4, B8) + 20 m bands (B5–B7 red-edge, B8A, B11/B12 SWIR) plus five normalised indices, each diagnostic of vegetation or water:

• NDVI = (B8−B4)/(B8+B4) — greenness/vigour [10]
• NDRE = (B8−B5)/(B8+B5) — red-edge chlorophyll [11]
• NDWI = (B3−B8)/(B3+B8) — open water [12]
• MNDWI = (B3−B11)/(B3+B11) — modified water index [13]
• NDMI = (B8−B11)/(B8+B11) — canopy moisture [14]

Composites are annual cloud-masked medians (scenes with <40 % cloud, SCL cloud / shadow / cirrus pixels removed).

2.3 Biomass and carbon — the satellite pathway (report §4)

“ETH 10 m canopy height (2020) → AGB via the Simard 2019 mangrove allometry gives a reference biomass; a Sentinel-2 spectral-index model extends it to all years.”

This sentence describes a three-step chain (08_biomass.py):

1. Reference canopy height. We take the ETH Global Canopy Height map (Lang et al. 2023 [4]) for 2020 — a 10 m product trained on GEDI space-borne LiDAR footprints and Sentinel-2 imagery — clipped to the AOI (AOI range 4–18 m).

2. Height → above-ground biomass (AGB) via a mangrove allometric power law of the form used by Simard et al. (2019) [15]:

   AGB [Mg ha⁻¹] = 10.44 · H^0.874        (H = canopy height in m)

   Simard et al. produced the global mangrove canopy-height and biomass maps by linking field-measured biomass–height allometry to remotely-sensed canopy height; their region-specific models carry RMSE of 54–103 Mg ha⁻¹. The single power law used here is an approximation in that framework (see Limitations). Applied to the ETH height it yields a reference per-pixel AGB for 2020.

3. Extend to all years with a spectral model. ETH height is single-epoch (2020 only), so to obtain a time series (2017–2025) we fit an empirical multiple linear regression from Sentinel-2 spectral predictors to the 2020 reference AGB, over mangrove pixels:

   AGB ≈ f(NDVI, NDRE, NDMI, B8, B11)     (ordinary least squares; training R² ≈ 0.34)

   The fitted model is then applied to each year’s Sentinel-2 composite, giving an AGB map per year. (The modest R² reflects spectral saturation of optical indices against biomass — a well-known limitation [16] — and is the main reason the LiDAR-direct estimate in §2.5 is more reliable.)

Carbon and CO₂. Per-cell AGB is converted to carbon and CO₂-equivalent with IPCC / blue-carbon defaults:

C_above = AGB · 0.451                         (carbon fraction, Hamilton & Friess 2018 [17])
C_total = AGB · (1 + 0.49) · 0.451            (+ below-ground biomass, root:shoot 0.49 [18,19])
CO₂e    = C_total · 44/12                      (molecular-mass ratio)

The carbon-sequestration rate is the slope of total carbon vs year (ordinary least squares; the report shows ≈ 9.7 Mg C yr⁻¹, trend R² ≈ 0.92).

2.4 Drone orthophoto and high-resolution clustering (report §5)

The 114 DJI Zenmuse L1 images are processed into a 2.5 cm orthophoto by OpenDroneMap [6] — Structure-from-Motion + multi-view stereo + orthorectification, georeferenced from the image RTK-GPS EXIF. Because the ortho is RGB-only and very high resolution, it is clustered with HDBSCAN [20,21], a density-based algorithm that (unlike k-means) finds the cohesive green-canopy density mode and labels ambiguous/mixed pixels as noise. Features are R, G, B plus two RGB indices: ExG = 2g−r−b (excess green [22]) and VARI = (g−r)/(g+r−b) [23]. The mangrove cluster is anchored to the polygons; the brightest non-green cluster (largest cell-count × brightness) is labelled bare ground (tidal flat).

2.5 LiDAR processing — validation and the final biomass (report §6–7)

The L1 cloud is binned to the 10 m satellite grid (for validation) and to the 0.5 m ortho grid (for the final biomass). Key steps:

• Canopy height model (CHM). Per cell, canopy top = 95th-percentile point height; ground is taken from the cluster-defined bare-ground cells (low percentile of returns) and interpolated under the closed canopy (linear, nearest-fill) — the standard way to recover ground beneath dense mangrove where no laser returns reach the floor [24]. CHM = canopy top − interpolated ground.
• Tree density. Individual tree tops are detected as local maxima of the CHM (a maximum filter with ~2.5 m minimum spacing, height ≥ 2 m), a standard LiDAR crown-detection approach [25]; density = tops / mangrove area.
• Biomass. The same Simard power law is applied to the LiDAR-measured CHM, computed side by side over two masks (the drone HDBSCAN mask and the AlphaEarth mask), sharing one ground surface.

2.6 Validation metrics

• Agreement % — fraction of co-located pixels with the same class.
• IoU (Jaccard) [26] — intersection ÷ union of the two mangrove masks.
• Cohen’s κ [27] — chance-corrected agreement (κ > 0.6 = substantial).
• RMSE / bias / Pearson r — for continuous comparisons (height, AGB).

3. Figure-by-figure guide

Each figure is generated by the script in parentheses and embedded in the report as a PNG.

1. Coverage time-series (04_visualize.py) — mangrove area (ha) per year, 2017–2025, from the AlphaEarth clustering. Area = (mangrove-cluster pixel count) × 100 m² (each 10 m UTM cell = 0.01 ha). Shows the rise 1.17 → 2.37 ha.

2. Change map (04_visualize.py) — the per-pixel growth / stable / decay classes (§2.1) drawn over the latest Sentinel-2 true-colour composite. Green = growth, blue = stable mangrove, red = decay (none here).

3. Yearly overlays (04_visualize.py) — one panel per year: the cloud-masked Sentinel-2 annual median true-colour image (B4/B3/B2) with the detected mangrove mask drawn in yellow. This is the visual check that the cluster tracks the real green patch as it expands.

4. Comparison — coverage (07_compare_clusterings.py) — two coverage curves, the AlphaEarth (64-band) vs Sentinel-2 (15-band) clusterings, showing both independent feature sets reproduce the same expansion.

5. Comparison — agreement map (07_compare_clusterings.py) — for the latest common year, each pixel coloured by whether both methods, only one, or neither call it mangrove. Title carries agreement %, IoU and κ (≈ 94 %, 0.73, 0.81).

6. Biomass (08_biomass.py) — left: total AGB (Mg) and carbon stock (Mg C) per year on twin axes, with the fitted sequestration slope; right: the 2025 per-pixel AGB map (Mg ha⁻¹). Built from the satellite pathway in §2.3.

7. Ortho cluster (10_ortho_cluster.py) — left: the drone orthophoto over the AOI; right: the same with the HDBSCAN mangrove cluster (yellow). Confirms the cm-scale delineation of the dense canopy.

8. LiDAR validation (09_lidar_validate.py) — four panels: (a) the 10 m LiDAR CHM; (b) LiDAR vs AlphaEarth mangrove-mask agreement; (c) scatter of LiDAR CHM vs ETH satellite height; (d) scatter of LiDAR AGB vs satellite AGB. Demonstrates strong extent agreement (κ up to 0.68) but weak per-pixel structural correlation (r ≈ 0.27–0.38) — satellite maps where, LiDAR resolves structure.

9. Final integrated biomass (12_lidar_biomass.py) — four panels: (a) the LiDAR CHM with detected tree tops; (b) the two mangrove masks compared (ortho-HDBSCAN vs AlphaEarth); (c) AGB over the ortho mask; (d) a bar chart of carbon stock by method/mask (satellite-spectral, global-baseline LiDAR, and the two final LiDAR×cluster estimates).

Interactive elements (not static figures)

• Map — Esri basemap with the drone orthophoto (continuous image) and the AlphaEarth / change / LiDAR-CHM layers drawn as exact 10 × 10 m cells (real per-pixel polygons, hover to read the value — no interpolation).
• 3-D viewer — a Three.js point cloud (200 k base + 600 k on demand) coloured by RGB, height, or the HDBSCAN cluster (green = mangrove, tan = ground).

4. Comparability with existing research

• Mangrove above-ground biomass. Global mangrove AGB spans roughly 10–500 Mg ha⁻¹, with mature stands typically 100–300 Mg ha⁻¹ [28,29]. Our estimates are at the low end — satellite-spectral ≈ 65–70 Mg ha⁻¹ and LiDAR-direct ≈ 22–29 Mg ha⁻¹ — which is expected and consistent with a young, short, expanding stand (LiDAR mean canopy height 2.6–3.2 m, vs the ~8 m global mangrove mean [15]). The height-allometry/spectral methods over-estimate relative to the direct LiDAR measurement precisely because they assume taller canopy; this is the headline integration finding.

• Canopy height. The LiDAR mean (2.6–3.2 m, max 15–18 m) and the ETH product (4–18 m) bracket each other; both are below mature mangrove canopy, again consistent with an establishing northern-limit site.

• Mangrove expansion. The Ryukyu Islands sit near the northern global range limit of mangroves, where poleward expansion and establishment are documented [30,31]. A ~doubling of cover over 2017–2025 is plausible for a young / restored site and is corroborated independently by the AlphaEarth and Sentinel-2 clusterings (κ ≈ 0.81) and by the drone/LiDAR snapshot.

• Blue carbon. Mangroves are among the most carbon-dense tropical forests [32]; our carbon-stock magnitudes (tens of Mg C over ~2–3 ha) are small only because the stand is small and young.

5. Limitations & uncertainty

• Allometric coefficients. The single power law AGB = 10.44·H^0.874 is an approximation of the region-specific Simard et al. (2019) framework [15], which itself carries 54–103 Mg ha⁻¹ RMSE; mangrove allometry is also species- and wood-density-dependent [18]. Field plots would be needed to calibrate locally.
• Spectral biomass model. Training R² ≈ 0.34 — optical indices saturate against biomass [16]; the per-year satellite AGB is therefore a coarse extrapolation, validated (and shown to over-estimate) by the LiDAR.
• Single-epoch height. ETH canopy height is 2020 only; the time series leans on the spectral model for inter-annual change.
• Unsupervised clusters. Cluster identity is anchored to the prior polygons, not field labels; coverage is sensitive to k and the 0.5 change threshold.
• 2017 Sentinel-2 L2A is unavailable over Japan, so the S2 record spans 2018–2025 while AlphaEarth covers 2017–2025.

6. Reproducibility

All numbers and figures are produced by the scripts in this folder (change_analysis/); see the main README.md for the pipeline and run_all.py. The orthophoto is produced separately by ../../Hiyagon_Orthophoto/run_odm_hiyagon.py. Inputs are public (Earth Engine, ESA) except the proprietary DJI L1 survey.

References

1. Brown, C. F. et al. (2025). AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. Google DeepMind. https://deepmind.google/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/
2. Google Earth Engine — Satellite Embedding V1 (Annual) data catalogue. https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL
3. Drusch, M. et al. (2012). Sentinel-2: ESA’s optical high-resolution mission for GMES operational services. Remote Sensing of Environment 120, 25–36.
4. Lang, N., Jetz, W., Schindler, K. & Wegner, J. D. (2023). A high-resolution canopy height model of the Earth. Nature Ecology & Evolution 7 (2023); preprint arXiv:2204.08322. GEE asset users/nlang/ETH_GlobalCanopyHeight_2020_10m_v1.
5. DJI. Zenmuse L1 product specifications (Livox LiDAR + IMU + 20 MP RGB).
6. OpenDroneMap Authors. ODM — open-source photogrammetry toolkit. https://www.opendronemap.org
7. Lloyd, S. P. (1982). Least squares quantization in PCM. IEEE Trans. Inf. Theory 28, 129–137. (k-means; orig. MacQueen 1967.)
8. Pedregosa, F. et al. (2011). scikit-learn: Machine learning in Python. JMLR 12, 2825–2830.
9. Rousseeuw, P. J. (1987). Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J. Comput. Appl. Math. 20, 53–65.
10. Rouse, J. W. et al. (1974). Monitoring vegetation systems in the Great Plains with ERTS. NASA SP-351. (NDVI.)
11. Gitelson, A. & Merzlyak, M. N. (1994). Spectral reflectance changes associated with autumn senescence… red-edge. J. Plant Physiol. 143, 286–292. (NDRE family.)
12. McFeeters, S. K. (1996). The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. Int. J. Remote Sens. 17, 1425–1432.
13. Xu, H. (2006). Modification of normalised difference water index (MNDWI) to enhance open water features… Int. J. Remote Sens. 27, 3025–3033.
14. Gao, B. C. (1996). NDWI — a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 58, 257–266. (NDMI/NDII family.)
15. Simard, M. et al. (2019). Mangrove canopy height globally related to precipitation, temperature and cyclone frequency. Nature Geoscience 12, 40–45.
16. Lu, D. (2006). The potential and challenge of remote sensing-based biomass estimation. Int. J. Remote Sens. 27, 1297–1328. (Spectral saturation.)
17. Hamilton, S. E. & Friess, D. A. (2018). Global carbon stocks and potential emissions due to mangrove deforestation from 2000 to 2012. Nature Climate Change 8, 240–244.
18. Komiyama, A., Ong, J. E. & Poungparn, S. (2008). Allometry, biomass, and productivity of mangrove forests: A review. Aquatic Botany 89, 128–137.
19. IPCC (2014). 2013 Supplement to the 2006 IPCC Guidelines for National Greenhouse Gas Inventories: Wetlands. (Below-ground biomass ratios.)
20. Campello, R. J. G. B., Moulavi, D. & Sander, J. (2013). Density-based clustering based on hierarchical density estimates. PAKDD, LNCS 7819.
21. McInnes, L., Healy, J. & Astels, S. (2017). hdbscan: Hierarchical density-based clustering. J. Open Source Software 2(11), 205.
22. Woebbecke, D. M. et al. (1995). Color indices for weed identification… (Excess Green, ExG.) Trans. ASAE 38, 259–269.
23. Gitelson, A. A. et al. (2002). Vegetation and soil lines in visible spectral space… (VARI.) Int. J. Remote Sens. 23, 2537–2562.
24. Popescu, S. C. et al. (2002). Measuring individual tree crown diameter with LiDAR and assessing its influence on estimating forest volume and biomass. Can. J. Remote Sens. 28, 564–577. (CHM from LiDAR.)
25. Popescu, S. C. & Wynne, R. H. (2004). Seeing the trees in the forest: using LiDAR and multispectral data fusion with local filtering and variable window size for estimating tree height. PE&RS 70, 589–604. (Local-maxima tree detection.)
26. Jaccard, P. (1912). The distribution of the flora in the alpine zone. New Phytologist 11, 37–50. (Jaccard / IoU.)
27. Cohen, J. (1960). A coefficient of agreement for nominal scales. Educ. Psychol. Meas. 20, 37–46. (Cohen’s κ.)
28. Hutchison, J. et al. (2014). Predicting global patterns in mangrove forest biomass. Conservation Letters 7, 233–240.
29. Rovai, A. S. et al. (2021). Macroecological patterns of forest structure and allometric scaling in mangrove forests. Global Ecology and Biogeography 30, 1000–1013.
30. Saintilan, N. et al. (2014). Mangrove expansion and salt marsh decline at mangrove poleward limits. Global Change Biology 20, 147–157.
31. Friess, D. A. et al. (2019). The state of the world’s mangrove forests: past, present, and future. Annual Review of Environment and Resources 44, 89–115.
32. Donato, D. C. et al. (2011). Mangroves among the most carbon-rich forests in the tropics. Nature Geoscience 4, 293–297.

Generated to accompany hiyagon_report.html. Figures and statistics are reproducible from the change_analysis/ pipeline. Citations [1], [4], [15] and [17] were verified against the primary sources; remaining references are the canonical primary sources for the named methods and should be consulted directly before quoting exact coefficients.