Application of the Kushiro Framework to the Chao Phraya Basin (2011 Flood) Without LiDAR

Author

Waruth

Published

March 23, 2026

This document presents a comprehensive adaptation of the Kushiro Framework for channel network extraction and validation to the Chao Phraya Basin’s 2011 flood event. The key innovation lies in transferring a methodology originally designed for high-resolution LiDAR data in a natural wetland to a data-limited, human-modified river basin using multi-source DEMs and historical flood extent data.

Key Adaptations:

DEM source diversity replaces ground filtering method diversity Historical flood inundation maps (2011) replace contemporary satellite imagery Dual-validation framework maintained: internal consistency (Level 1) + external plausibility (Level 2) Integration with RRI hydrological model for physically-based flood simulation

Main Challenges in Transferring Methodology

First, it’s essential to understand the fundamental differences between the two areas, as they directly affect how the methodology must be adapted:

Issue	Kushiro Wetland	Chao Phraya Central Plain
Comparison of Study Areas
Kushiro Wetland vs. Chao Phraya Basin
Topography Data	UAV-LiDAR (4.76 pts/m²)	SRTM / ALOS AW3D30 (30 m)
DTM Resolution	5 m	30–90 m
Vertical Accuracy	~0.05–0.10 m	~1–5 m (SRTM RMSE)
Channel Characteristics	Natural, 1st–3rd order	Modified + Irrigation canals (Khlongs)
Vegetation	Reed, Sedge, Alder	Rice paddy, Urban, Heterogeneous vegetation
Area Size	0.91 km²	~20,000 km² (2011 flood extent)
Average Slope	< 0.5°	< 0.1° (flatter than Kushiro)

The Vertical Accuracy Mismatch Problem

The biggest problem is the vertical accuracy mismatch: the available DEM’s vertical accuracy is coarser than the topographic relief that needs to be classified. This is the opposite of Kushiro, where LiDAR provided resolution far exceeding channel depth.

Available Topography Data Options

Since LiDAR is unavailable, multiple DEM sources must be used instead of comparing ground filtering algorithms as in the Kushiro study. Different DEM sources essentially serve the same function as different ground filtering methods in the context of systematic workflow evaluation.

SRTM v3 (30 m) has global coverage and is widely used, but suffers from vegetation bias because C-band radar is affected by tree canopy, meaning the resulting surface isn’t true bare earth. In mangrove or palm plantation areas, bias can reach 2–5 m.

ALOS AW3D30 (30 m) uses optical stereo from PRISM sensors, producing elevations closer to DSM than DTM. However, in open areas like rice paddies in the Chao Phraya floodplain, the difference between DSM and DTM is minimal, making ALOS quite accurate in this context.

Copernicus DEM GLO-30 (30 m), created from TanDEM-X interferometry, offers significantly better vertical accuracy than SRTM (RMSE ~1 m in flat areas) and is considered the best free DEM currently available.

TanDEM-X 12 m (commercial/academic access): If academic licensing can be obtained, this provides significantly higher resolution, ideal for main channel delineation of the Chao Phraya River.

DEM Source	Resolution	Acquisition Method	Vertical Accuracy	Key Characteristics	Access
Available DEM Sources for Chao Phraya Basin
Functional equivalents to ground filtering algorithms in Kushiro study
SRTM v3	30 m (~1 arc-sec)	C-band InSAR (2000)	~3-5 m (flat) ~6-9 m (vegetated)	Global coverage, widely used; Vegetation bias (C-band affected by canopy); Bias 2-5 m in mangrove/palm areas	Free (USGS)
ALOS AW3D30	30 m	Optical stereo PRISM (2006-2011)	~2-4 m (open) DSM ≈ DTM in paddies	Optical stereo produces DSM; Minimal DSM-DTM difference in open areas; Accurate for Chao Phraya floodplain	Free (JAXA)
Copernicus DEM GLO-30	30 m	X-band InSAR (2010-2015)	~1-2 m (best available)	Best free DEM currently available; Significantly better than SRTM; Ideal for primary analysis	Free (Copernicus)
TanDEM-X 12m	12 m (0.4 arc-sec)	X-band InSAR (2010-2015)	~2 m (flat) Highest resolution	Academic/commercial access required; Ideal for main channel delineation; Highest detail available	Academic license

Re-designing the Workflow for Available Data

Conceptual Mapping: Kushiro → Chao Phraya

The Kushiro framework remains applicable with adjusted component mapping:

Kushiro Component	Function	Chao Phraya Equivalent	Rationale
Conceptual Framework Mapping
Kushiro (UAV-LiDAR) → Chao Phraya (Multi-source DEM)
Ground Filtering (PMF/CSF/MCC)	Remove vegetation, extract bare earth	DEM Source (SRTM/ALOS/Copernicus/TanDEM-X)	Different acquisition methods = different surface representations
DTM Interpolation (IDW/TIN/KRG)	Convert point cloud to raster surface	DEM Preprocessing (Hydrological Conditioning)	Reconditioning needed for human-modified landscape
Sink Filling (Wang/Planchon)	Remove spurious depressions	Sink Filling (Wang/Planchon) [unchanged]	Same algorithm applies to raster DEMs
Flow Direction (D8/D-Inf)	Determine water flow paths	Flow Direction (D8/D-Inf) [unchanged]	Same algorithm applies regardless of source
Validation: Sentinel-2	External reality check	Validation: 2011 Flood Inundation Maps	Historical flood extent replaces contemporary imagery

Additional Workflow Components for Chao Phraya

Due to greater area complexity compared to Kushiro, additional components are needed for the human-modified drainage network. This includes incorporating canal networks from OpenStreetMap or the Royal Irrigation Department to “burn in” known channels into the DEM before extraction—a technique called Stream Burning or Agreement Burning. This increases flow accumulation sensitivity in canal areas.

Additional DEM reconditioning is also necessary because the Chao Phraya floodplain has numerous roads and embankments crossing waterways, creating artificial barriers in the DEM. Breach algorithms must be used to “cut through” these ridges before sink filling.

Component	Why Needed	Data Source	Method	Impact on Results
Additional Workflow Components for Chao Phraya
Handling human-modified landscape complexity
Stream Burning / Agreement Burning	Human-created canal networks not always visible in DEM	OpenStreetMap waterway data; Royal Irrigation Department canal maps	Burn known channels into DEM before extraction; Increase flow accumulation in canal areas; Ensure routing follows known waterways	Ensures extracted network includes anthropogenic channels; Improves alignment with actual drainage
DEM Reconditioning for Infrastructure	Roads and embankments create artificial barriers in DEM	OpenStreetMap road network; Google Earth imagery for visual verification	Apply breach algorithms to 'cut through' ridges; Identify road crossings and embankments; Remove artificial flow barriers	Prevents fragmentation of drainage network; Reduces unrealistic flow accumulation patterns
Canal Network Integration	Extensive irrigation system central to drainage pattern	Royal Irrigation Department (RID); OpenStreetMap (waterway=canal tags)	Vector-to-raster conversion; Combine with natural drainage network; Validate flow connectivity	More accurate representation of actual flow paths; Better prediction of flood routing

Workflow Combinations Matrix

Component	Options (Count)	Details	Kushiro Equivalent
Workflow Components for Systematic Evaluation
Total combinations: 4 × 2 × 2 × 3 = 48 workflows
DEM Source	4 sources	SRTM, ALOS, Copernicus, TanDEM-X	Ground filtering (PMF/CSF/MCC)
Sink Filling	2 methods	Wang & Liu (2006), Planchon & Darboux (2002)	Same algorithm
Flow Direction	2 methods	D8 (single direction), D-Infinity (multiple direction)	Same algorithm
Flow Accumulation Threshold	3 thresholds	Based on drainage area: 1 km², 2.5 km², 5 km²	Area-based threshold (5,000 m²)

This includes:

Incorporating canal networks from OpenStreetMap or the Royal Irrigation Department to “burn in” known channels into the DEM before extraction—a technique called Stream Burning or Agreement Burning. This increases flow accumulation sensitivity in canal areas.
Additional DEM reconditioning is also necessary because the Chao Phraya floodplain has numerous roads and embankments crossing waterways, creating artificial barriers in the DEM. Breach algorithms must be used to “cut through” these ridges before sink filling.

Using 2011 Flood Data for External Validation

This is the most important strength of this application, as the 2011 flood has abundant remote sensing data that can replace Sentinel-2 from the Kushiro study.

Available Flood Inundation Data

Data Source	Spatial Resolution	Temporal Coverage	Key Advantage	Use in Validation
Available 2011 Flood Inundation Data Sources
Remote sensing data for external validation
MODIS Terra/Aqua	250–500 m	Daily throughout flood period	Daily data for tracking temporal evolution NDWI/MNDWI can detect surface water quickly Ideal for flood front progression analysis	Track flood propagation timing Validate temporal consistency of network
Landsat 5/7	30 m	16-day revisit (some cloud gaps)	Better spatial resolution than MODIS Band 5 (SWIR) sensitive to surface water Can overlay with extracted channel networks	High-resolution channel-flood alignment Primary validation dataset
ALOS-2 PALSAR / RADARSAT-2 (SAR)	10–25 m	Continuous (cloud-penetrating)	SAR penetrates clouds Continuous data during monsoon period More accurate flood extent than optical NASA/Dartmouth Flood Observatory compiled maps	Ground truth for cloud-covered periods Validate flood corridor extent
JAXA GCOM-W AMSR2	~10 km	Daily	Daily flood inundation data Useful for hydrological connectivity validation Low resolution limits detailed comparison	Basin-scale connectivity check Supplementary validation

MODIS Terra/Aqua (250–500 m) provides daily data throughout the flood period. NDWI or Modified NDWI can detect surface water quickly, ideal for tracking temporal evolution of the flood front.

Landsat 5/7 (30 m) offers much better spatial resolution than MODIS, especially Band 5 (SWIR), which is sensitive to surface water. Landsat results can be overlaid with extracted channel networks to validate whether predicted channels correspond with actual flood paths.

ALOS-2 PALSAR / RADARSAT-2 (SAR): SAR penetrates clouds, providing continuous data throughout the monsoon flood period with more accurate flood extent than optical sensors during rain. NASA Flood Observatory and Dartmouth Flood Observatory have already compiled SAR flood maps from this event.

JAXA GCOM-W AMSR2 provides daily flood inundation data. Though spatial resolution is relatively low (~10 km), it’s useful for validating that the derived channel network has reasonable hydrological connectivity.

Two-Level Validation Framework Design (Analogous to Kushiro)

Validation Level	Metric	Formula	Data Required	Purpose
Two-Level Validation Framework
Analogous to Kushiro study structure, adapted for 2011 flood context
Level 1: Pairwise Statistical Consensus	Intersection over Union (IoU)	IoU = TP / (TP + FP + FN)	Multiple channel extractions from workflow combinations	Measure agreement between different workflow outputs (internal consistency)
Level 1: Pairwise Statistical Consensus	F1-Score	F1 = 2TP / (2TP + FP + FN)	Same as IoU	Alternative metric for binary classification performance
Level 2: External Plausibility Validation	Channel-Flood Correspondence	CFC = Channel pixels within flood extent / Total channel pixels	Extracted channel network + 2011 flood extent map	Check if predicted channels align with actual flood corridors
Level 2: External Plausibility Validation	Flood Routing Plausibility	FRP = Flood progression matches network flow direction	Channel network + temporal flood progression (MODIS)	Verify that flood propagation follows extracted network topology
Level 2: External Plausibility Validation	Temporal Consistency Index	TCI = Correlation between flow accumulation and flood timing	Flow accumulation raster + multi-date flood maps	Validate that upstream areas flood before downstream areas

Level 1: Pairwise Statistical Consensus remains the same as before—comparing channel network outputs from different workflow combinations using the IoU metric from the Kushiro study:

\[ IoU = \frac{𝑇P}{TP+FP+FN} \tag{1}\]

Level 2: External Plausibility Validation uses 2011 flood inundation extent instead of Sentinel-2, based on the principle that “a good channel network should predict flow paths consistent with actual flood extent.”

This calculates spatial overlap between predicted channels and observed flood corridors:

\[ ChannelFlood Correspondence = \frac{Channel Pixels Within Flood Extent}{Total Channel Pixels} \]

A new metric appropriate for the 2011 flood context is added: Flood Routing Plausibility, which checks whether the extracted network can explain the propagation of the flood front from upstream to downstream.

Clear Research Novelty and Contribution

The Kushiro study’s novelty lies in its workflow evaluation framework for UAV-LiDAR in wetland areas. To publish Chao Phraya work, distinct novelty is needed:

Novelty Aspect	Kushiro Baseline	Chao Phraya Innovation	Scientific Contribution
Research Novelty and Contribution
Distinct contributions beyond Kushiro framework
1. Framework Transferability across Data Availability Contexts	Dual-validation framework applied to high-resolution LiDAR in natural wetland	Same framework applied to multi-source DEMs in data-poor context DEM source diversity replaces ground filtering diversity Demonstrates methodological robustness across data types	Methodological framework proves applicable beyond original context Provides template for other data-limited regions Validates dual-validation concept as generalizable
2. Flood Inundation as Hydrological Validation	Contemporary Sentinel-2 imagery for visual validation of channel presence	Historical flood extent (2011) as proxy for network plausibility Temporal flood progression validates network topology Systematic use of flood data rarely done in channel extraction	Establishes flood inundation as valid hydrological proxy Addresses temporal gap between DEM and validation data Creates methodology for post-event validation
3. Human-Modified vs. Natural Channel Networks	Natural channel network (1st-3rd order streams) in pristine wetland	Extensive human-created canal network (khlongs) Complex mix of natural channels and irrigation infrastructure Evaluate automated extraction capability for both types	Addresses understudied problem of mixed natural-anthropogenic networks Relevant for heavily modified river basins globally Quantifies extraction performance by channel type

Novelty 1: Framework Transferability across Data Availability Contexts demonstrates that the dual-validation framework from Kushiro can transfer to more data-poor situations, where DEM source diversity replaces ground filtering diversity and flood inundation maps replace contemporaneous satellite imagery.

Novelty 2: Flood Inundation as Hydrological Validation uses historical flood extent as a proxy for network plausibility—something few studies have done systematically. The challenge is explaining the temporal gap between DEM acquisition and the flood event (DEM may be from a different year than 2011).

Novelty 3: Human-Modified vs. Natural Channel Networks: The Chao Phraya has an extensive human-created canal network, making flow routing far more complex than Kushiro. Evaluating how well automated extraction can recover both natural channels and anthropogenic canals is scientifically valuable.

Cautions and Limitations

Limitation	Problem Description	Impact on Results	Mitigation Strategy
Cautions and Limitations
Critical challenges requiring careful handling
Vertical Resolution Mismatch	In Chao Phraya floodplain, elevation variation between channels and floodplains may be only 0.5–2 m This falls within SRTM's RMSE range (3-5 m) Creates high uncertainty in channel delineation from DEM	Channel detection uncertainty increases dramatically False positives and false negatives both likely Extracted network may miss shallow channels May incorrectly identify non-channel features	Use multi-source DEM consensus to reduce individual source bias Perform comprehensive sensitivity analysis Quantify uncertainty ranges for all results Clearly report confidence levels by terrain type
Temporal Inconsistency	DEM captured before 2011 (SRTM: 2000, ALOS: 2006-2011) Flood event occurred in 2011 Canals may have been dug or rice fields filled during interval Landscape changes not reflected in DEM	Validation may show poor correspondence in areas of landscape change Extracted network may not match 2011 flood paths if canals changed Reduces reliability of flood-based validation	Document all known landscape changes between DEM date and 2011 Use multiple temporal DEM sources if available Focus validation on stable landscape areas Explicitly acknowledge temporal mismatch in limitations
Scale Dependency	Flow accumulation threshold of 5,000 m² used in Kushiro at 5 m resolution cannot be directly applied 30 m resolution in Chao Phraya requires different thresholds Drainage area to classify must be recalculated	Inappropriate thresholds lead to over/under-segmentation Network density may not match reality Comparisons with Kushiro results require careful normalization	Calculate new thresholds based on drainage area analysis Perform scale-dependent validation Test multiple threshold values systematically Report results as functions of scale parameters

The biggest problem is vertical resolution mismatch: in the Chao Phraya floodplain, elevation variation between channels and floodplains may be only 0.5–2 m, which falls within SRTM’s RMSE range. This creates high uncertainty in channel delineation from DEM, requiring clear sensitivity analysis and uncertainty quantification.

Another issue is temporal inconsistency between the DEM (captured before 2011) and the flood event (2011), especially in areas where canals were dug or rice fields filled during that interval. This must be considered and documented in the limitations.

Finally, scale dependency is critical because the flow accumulation threshold of 5,000 m² used in Kushiro at 5 m resolution cannot be directly applied to 30 m resolution in Chao Phraya. New thresholds must be calculated based on the drainage area to be classified.

Integration with RRI Hydrological Model

Complete Workflow Diagram

Complete workflow from data inputs through validation to final output

RRI Model Configuration

The Rainfall-Runoff-Inundation (RRI) model is a 2D physically-based distributed hydrological model that simultaneously solves:

Rainfall-runoff on hillslope grid cells
1D river flow in channel network
2D overland flow on floodplains
Lateral exchange between river and floodplain

Parameter	Value	Justification
RRI Model Configuration
Setup for 2011 Chao Phraya flood simulation
Model Domain	Chao Phraya Basin	Covers 2011 flood extent (~20,000 km²)
Grid Resolution	250 m × 250 m	Balance between accuracy and computational cost
Simulation Period	July 1 - December 31, 2011 (184 days)	Captures entire 2011 monsoon flood event
Time Step	60 seconds (adaptive)	Ensures numerical stability (CFL condition)
Channel Network	From optimal Kushiro workflow	Validated drainage network from framework
Rainfall Input	GPM IMERG Final Run (30-min)	Best available gridded precipitation data
Boundary Conditions	C.2 station discharge (upstream)	Chiang Mai station provides upstream inflow
Initial Conditions	Dry condition (water depth = 0)	Monsoon onset marks simulation start
Output Frequency	Daily inundation maps + hourly hydrographs	Daily for validation, hourly for dynamics

--- title: "Application of the Kushiro Framework to the Chao Phraya Basin (2011 Flood) Without LiDAR" author: "Waruth" date: today format: html: toc: true toc-depth: 3 code-fold: true code-tools: true theme: cosmo fig-width: 10 fig-height: 6 --- ```{r setup} #| include: false knitr::opts_chunk$set( echo = FALSE, warning = FALSE, message = FALSE, fig.width = 10, fig.height = 6, dpi = 300 ) #| include: false # Load required packages library(tidyverse) library(knitr) library(kableExtra) library(gt) library(gtExtras) library(ggplot2) library(patchwork) library(DiagrammeR) library(scales) library(viridis) # Helper function for dual-format tables create_table <- function(data, title = NULL, subtitle = NULL, ...) { if (knitr::is_latex_output()) { # PDF output using kableExtra kbl <- data %>% kbl(booktabs = TRUE, caption = if(!is.null(title)) paste0(title, if(!is.null(subtitle)) paste0(": ", subtitle) else ""), longtable = TRUE, ...) %>% kable_styling( latex_options = c("striped", "scale_down", "HOLD_position"), font_size = 9, full_width = FALSE ) %>% row_spec(0, bold = TRUE, color = "white", background = "#4472C4") return(kbl) } else { # HTML output using gt gt_table <- data %>% gt() %>% tab_options( table.font.size = 11, heading.title.font.size = 14, heading.subtitle.font.size = 12 ) if (!is.null(title)) { gt_table <- gt_table %>% tab_header( title = title, subtitle = subtitle ) } return(gt_table) } } ``` This document presents a comprehensive adaptation of the Kushiro Framework for channel network extraction and validation to the Chao Phraya Basin's 2011 flood event. The key innovation lies in transferring a methodology originally designed for high-resolution LiDAR data in a natural wetland to a data-limited, human-modified river basin using multi-source DEMs and historical flood extent data. Key Adaptations: DEM source diversity replaces ground filtering method diversity Historical flood inundation maps (2011) replace contemporary satellite imagery Dual-validation framework maintained: internal consistency (Level 1) + external plausibility (Level 2) Integration with RRI hydrological model for physically-based flood simulation ## Main Challenges in Transferring Methodology First, it's essential to understand the fundamental differences between the two areas, as they directly affect how the methodology must be adapted: ```{r} #| echo: false # Create comparison dataframe comparison_df <- tibble( Issue = c( "Topography Data", "DTM Resolution", "Vertical Accuracy", "Channel Characteristics", "Vegetation", "Area Size", "Average Slope" ), `Kushiro Wetland` = c( "UAV-LiDAR (4.76 pts/m²)", "5 m", "~0.05–0.10 m", "Natural, 1st–3rd order", "Reed, Sedge, Alder", "0.91 km²", "< 0.5°" ), `Chao Phraya Central Plain` = c( "SRTM / ALOS AW3D30 (30 m)", "30–90 m", "~1–5 m (SRTM RMSE)", "Modified + Irrigation canals (Khlongs)", "Rice paddy, Urban, Heterogeneous vegetation", "~20,000 km² (2011 flood extent)", "< 0.1° (flatter than Kushiro)" ) ) # Display table create_table( comparison_df, title = "Comparison of Study Areas", subtitle = "Kushiro Wetland vs. Chao Phraya Basin" ) %>% {if(knitr::is_latex_output()) { column_spec(., 1, bold = TRUE, width = "3cm") %>% column_spec(., 2, width = "4cm") %>% column_spec(., 3, width = "6cm") } else { tab_style(., style = cell_fill(color = "#f0f0f0"), locations = cells_body(columns = Issue) ) %>% tab_style(., style = cell_text(weight = "bold"), locations = cells_body(columns = Issue) ) %>% cols_width( Issue ~ px(150), `Kushiro Wetland` ~ px(150), `Chao Phraya Central Plain` ~ px(150) ) }} ``` ## The Vertical Accuracy Mismatch Problem The biggest problem is the vertical accuracy mismatch: the available DEM's vertical accuracy is coarser than the topographic relief that needs to be classified. This is the opposite of Kushiro, where LiDAR provided resolution far exceeding channel depth. ```{r} #| eval: false #| fig-cap: Vertical accuracy mismatch between channel depth and DEM accuracy #| fig-height: 5 #| include: false # Create visualization of the accuracy mismatch problem challenge_data <- tibble( Feature = rep(c("Channel Depth", "DEM Accuracy"), 2), Location = c("Kushiro", "Kushiro", "Chao Phraya", "Chao Phraya"), Lower = c(0.5, 0.05, 0.5, 1), Upper = c(2, 0.10, 2, 5), Midpoint = c(1.25, 0.075, 1.25, 3) ) ggplot(challenge_data, aes(x = Location, y = Midpoint, color = Feature)) + geom_point(size = 4) + geom_errorbar(aes(ymin = Lower, ymax = Upper), width = 0.2, size = 1) + geom_hline(yintercept = 1.25, linetype = "dashed", color = "gray50", alpha = 0.7) + annotate("rect", xmin = 0.5, xmax = 1.5, ymin = 0.5, ymax = 2, alpha = 0.1, fill = "green") + annotate("text", x = 1, y = 2.3, label = "LiDAR accuracy\nexceeds channel depth", size = 3.5, fontface = "bold", color = "darkgreen") + annotate("rect", xmin = 1.5, xmax = 2.5, ymin = 0.5, ymax = 2, alpha = 0.1, fill = "red") + annotate("text", x = 2, y = 2.3, label = "DEM accuracy\ncoarser than relief", size = 3.5, fontface = "bold", color = "darkred") + scale_color_manual(values = c("Channel Depth" = "#F18F01", "DEM Accuracy" = "#6A4C93")) + labs( title = "The Vertical Accuracy Mismatch Problem", subtitle = "Comparison of typical channel depth vs. DEM vertical accuracy", y = "Elevation (meters)", x = NULL, caption = "Green zone = favorable conditions; Red zone = challenging conditions" ) + theme_minimal(base_size = 12) + theme( legend.position = "top", legend.title = element_blank(), plot.title = element_text(face = "bold", size = 14), plot.caption = element_text(hjust = 0, face = "italic") ) ``` ## Available Topography Data Options Since LiDAR is unavailable, multiple DEM sources must be used instead of comparing ground filtering algorithms as in the Kushiro study. Different DEM sources essentially serve the same function as different ground filtering methods in the context of systematic workflow evaluation. **SRTM v3 (30 m)** has global coverage and is widely used, but suffers from vegetation bias because C-band radar is affected by tree canopy, meaning the resulting surface isn't true bare earth. In mangrove or palm plantation areas, bias can reach 2–5 m. **ALOS AW3D30 (30 m)** uses optical stereo from PRISM sensors, producing elevations closer to DSM than DTM. However, in open areas like rice paddies in the Chao Phraya floodplain, the difference between DSM and DTM is minimal, making ALOS quite accurate in this context. **Copernicus DEM GLO-30 (30 m)**, created from TanDEM-X interferometry, offers significantly better vertical accuracy than SRTM (RMSE ~1 m in flat areas) and is considered the best free DEM currently available. **TanDEM-X 12 m (commercial/academic access)**: If academic licensing can be obtained, this provides significantly higher resolution, ideal for main channel delineation of the Chao Phraya River. ```{r echo=FALSE} # Create comprehensive DEM comparison table dem_sources <- tibble( `DEM Source` = c( "SRTM v3", "ALOS AW3D30", "Copernicus DEM GLO-30", "TanDEM-X 12m" ), `Resolution` = c( "30 m (~1 arc-sec)", "30 m", "30 m", "12 m (0.4 arc-sec)" ), `Acquisition Method` = c( "C-band InSAR (2000)", "Optical stereo PRISM (2006-2011)", "X-band InSAR (2010-2015)", "X-band InSAR (2010-2015)" ), `Vertical Accuracy` = c( "~3-5 m (flat)\n~6-9 m (vegetated)", "~2-4 m (open)\nDSM ≈ DTM in paddies", "~1-2 m\n(best available)", "~2 m (flat)\nHighest resolution" ), `Key Characteristics` = c( "Global coverage, widely used; Vegetation bias (C-band affected by canopy); Bias 2-5 m in mangrove/palm areas", "Optical stereo produces DSM; Minimal DSM-DTM difference in open areas; Accurate for Chao Phraya floodplain", "Best free DEM currently available; Significantly better than SRTM; Ideal for primary analysis", "Academic/commercial access required; Ideal for main channel delineation; Highest detail available" ), `Access` = c( "Free (USGS)", "Free (JAXA)", "Free (Copernicus)", "Academic license" ) ) create_table( dem_sources, title = "Available DEM Sources for Chao Phraya Basin", subtitle = "Functional equivalents to ground filtering algorithms in Kushiro study" ) %>% {if(knitr::is_latex_output()) { column_spec(., 1, bold = TRUE, width = "2.5cm") %>% column_spec(., 4, width = "2.5cm") %>% column_spec(., 5, width = "3cm") %>% row_spec(., 3, background = "#d4edda") # Highlight Copernicus } else { tab_style(., style = cell_fill(color = "#d4edda"), locations = cells_body(rows = 3) # Highlight Copernicus ) %>% tab_style(., style = cell_text(weight = "bold"), locations = cells_body(columns = `DEM Source`) ) %>% cols_width( `DEM Source` ~ px(120), Resolution ~ px(100), `Acquisition Method` ~ px(120), `Vertical Accuracy` ~ px(120), `Key Characteristics` ~ px(120), Access ~ px(100) ) }} ``` ## Re-designing the Workflow for Available Data ### Conceptual Mapping: Kushiro → Chao Phraya The Kushiro framework remains applicable with adjusted component mapping: ```{r echo=FALSE} # Create conceptual mapping table workflow_mapping <- tibble( `Kushiro Component` = c( "Ground Filtering (PMF/CSF/MCC)", "DTM Interpolation (IDW/TIN/KRG)", "Sink Filling (Wang/Planchon)", "Flow Direction (D8/D-Inf)", "Validation: Sentinel-2" ), `Function` = c( "Remove vegetation, extract bare earth", "Convert point cloud to raster surface", "Remove spurious depressions", "Determine water flow paths", "External reality check" ), `Chao Phraya Equivalent` = c( "DEM Source (SRTM/ALOS/Copernicus/TanDEM-X)", "DEM Preprocessing (Hydrological Conditioning)", "Sink Filling (Wang/Planchon) [unchanged]", "Flow Direction (D8/D-Inf) [unchanged]", "Validation: 2011 Flood Inundation Maps" ), `Rationale` = c( "Different acquisition methods = different surface representations", "Reconditioning needed for human-modified landscape", "Same algorithm applies to raster DEMs", "Same algorithm applies regardless of source", "Historical flood extent replaces contemporary imagery" ) ) create_table( workflow_mapping, title = "Conceptual Framework Mapping", subtitle = "Kushiro (UAV-LiDAR) → Chao Phraya (Multi-source DEM)" ) %>% {if(knitr::is_latex_output()) { column_spec(., 1, bold = TRUE, width = "3.5cm") %>% column_spec(., 2, width = "3cm") %>% column_spec(., 3, width = "3.5cm") %>% column_spec(., 4, width = "3cm") %>% row_spec(., c(1, 2, 5), background = "#e3f2fd") # Highlight changed components } else { tab_style(., style = cell_fill(color = "#e3f2fd"), locations = cells_body(rows = c(1, 2, 5)) # Components that change ) %>% tab_style(., style = cell_fill(color = "#fff3cd"), locations = cells_body(rows = c(3, 4)) # Components that remain ) %>% cols_width( `Kushiro Component` ~ px(150), Function ~ px(150), `Chao Phraya Equivalent` ~ px(150), Rationale ~ px(150) ) }} ``` ## Additional Workflow Components for Chao Phraya Due to greater area complexity compared to Kushiro, additional components are needed for the human-modified drainage network. This includes incorporating canal networks from OpenStreetMap or the Royal Irrigation Department to "burn in" known channels into the DEM before extraction—a technique called Stream Burning or Agreement Burning. This increases flow accumulation sensitivity in canal areas. Additional DEM reconditioning is also necessary because the Chao Phraya floodplain has numerous roads and embankments crossing waterways, creating artificial barriers in the DEM. Breach algorithms must be used to "cut through" these ridges before sink filling. ```{r echo=FALSE} # Additional workflow components additional_components <- tibble( Component = c( "Stream Burning / Agreement Burning", "DEM Reconditioning for Infrastructure", "Canal Network Integration" ), `Why Needed` = c( "Human-created canal networks not always visible in DEM", "Roads and embankments create artificial barriers in DEM", "Extensive irrigation system central to drainage pattern" ), `Data Source` = c( "OpenStreetMap waterway data; Royal Irrigation Department canal maps", "OpenStreetMap road network; Google Earth imagery for visual verification", "Royal Irrigation Department (RID); OpenStreetMap (waterway=canal tags)" ), `Method` = c( "Burn known channels into DEM before extraction; Increase flow accumulation in canal areas; Ensure routing follows known waterways", "Apply breach algorithms to 'cut through' ridges; Identify road crossings and embankments; Remove artificial flow barriers", "Vector-to-raster conversion; Combine with natural drainage network; Validate flow connectivity" ), `Impact on Results` = c( "Ensures extracted network includes anthropogenic channels; Improves alignment with actual drainage", "Prevents fragmentation of drainage network; Reduces unrealistic flow accumulation patterns", "More accurate representation of actual flow paths; Better prediction of flood routing" ) ) create_table( additional_components, title = "Additional Workflow Components for Chao Phraya", subtitle = "Handling human-modified landscape complexity" ) %>% {if(knitr::is_latex_output()) { column_spec(., 1, bold = TRUE, width = "3cm") %>% column_spec(., 2:5, width = "3cm") } else { tab_style(., style = cell_fill(color = "#fff3cd"), locations = cells_body(columns = Component) ) %>% tab_style(., style = cell_text(weight = "bold"), locations = cells_body(columns = Component) ) %>% cols_width( Component ~ px(150), `Why Needed` ~ px(150), `Data Source` ~ px(150), Method ~ px(150), `Impact on Results` ~ px(150) ) }} ``` ## Workflow Combinations Matrix ```{r echo=FALSE} # Create workflow components summary workflow_components <- tibble( Component = c( "DEM Source", "Sink Filling", "Flow Direction", "Flow Accumulation Threshold" ), `Options (Count)` = c( "4 sources", "2 methods", "2 methods", "3 thresholds" ), Details = c( "SRTM, ALOS, Copernicus, TanDEM-X", "Wang & Liu (2006), Planchon & Darboux (2002)", "D8 (single direction), D-Infinity (multiple direction)", "Based on drainage area: 1 km², 2.5 km², 5 km²" ), `Kushiro Equivalent` = c( "Ground filtering (PMF/CSF/MCC)", "Same algorithm", "Same algorithm", "Area-based threshold (5,000 m²)" ) ) create_table( workflow_components, title = "Workflow Components for Systematic Evaluation", subtitle = "Total combinations: 4 × 2 × 2 × 3 = 48 workflows" ) %>% {if(knitr::is_latex_output()) { column_spec(., 1, bold = TRUE, width = "3.5cm") %>% column_spec(., 2, width = "2.5cm") %>% column_spec(., 3, width = "4cm") %>% column_spec(., 4, width = "3cm") } else { tab_style(., style = cell_fill(color = "#f3e5f5"), locations = cells_body(columns = Component) ) %>% tab_style(., style = cell_text(weight = "bold"), locations = cells_body(columns = Component) ) %>% cols_width( Component ~ px(100), `Options (Count)` ~ px(80), Details ~ px(150), `Kushiro Equivalent` ~ px(120) ) }} ``` **This includes:** - Incorporating canal networks from OpenStreetMap or the Royal Irrigation Department to "burn in" known channels into the DEM before extraction—a technique called Stream Burning or Agreement Burning. This increases flow accumulation sensitivity in canal areas. - Additional DEM reconditioning is also necessary because the Chao Phraya floodplain has numerous roads and embankments crossing waterways, creating artificial barriers in the DEM. Breach algorithms must be used to "cut through" these ridges before sink filling. ## Using 2011 Flood Data for External Validation This is the most important strength of this application, as the 2011 flood has abundant remote sensing data that can replace Sentinel-2 from the Kushiro study. ### Available Flood Inundation Data ```{r} #| echo: false # Create flood data sources table flood_data <- tibble( `Data Source` = c( "MODIS Terra/Aqua", "Landsat 5/7", "ALOS-2 PALSAR / RADARSAT-2 (SAR)", "JAXA GCOM-W AMSR2" ), `Spatial Resolution` = c( "250–500 m", "30 m", "10–25 m", "~10 km" ), `Temporal Coverage` = c( "Daily throughout flood period", "16-day revisit (some cloud gaps)", "Continuous (cloud-penetrating)", "Daily" ), `Key Advantage` = c( "Daily data for tracking temporal evolution\nNDWI/MNDWI can detect surface water quickly\nIdeal for flood front progression analysis", "Better spatial resolution than MODIS\nBand 5 (SWIR) sensitive to surface water\nCan overlay with extracted channel networks", "SAR penetrates clouds\nContinuous data during monsoon period\nMore accurate flood extent than optical\nNASA/Dartmouth Flood Observatory compiled maps", "Daily flood inundation data\nUseful for hydrological connectivity validation\nLow resolution limits detailed comparison" ), `Use in Validation` = c( "Track flood propagation timing\nValidate temporal consistency of network", "High-resolution channel-flood alignment\nPrimary validation dataset", "Ground truth for cloud-covered periods\nValidate flood corridor extent", "Basin-scale connectivity check\nSupplementary validation" ) ) flood_data %>% gt() %>% tab_header( title = "Available 2011 Flood Inundation Data Sources", subtitle = "Remote sensing data for external validation" ) %>% tab_style( style = cell_text(weight = "bold"), locations = cells_column_labels() ) %>% tab_style( style = cell_fill(color = "white"), locations = cells_body( columns = `Data Source`, rows = c(2, 3) # Highlight best sources ) ) %>% cols_width( `Data Source` ~ px(100), `Spatial Resolution` ~ px(100), `Temporal Coverage` ~ px(100), `Key Advantage` ~ px(100), `Use in Validation` ~ px(100) ) %>% tab_style( style = cell_text(size = px(10)), locations = cells_body(columns = `Key Advantage`) ) %>% tab_options( table.font.size = 11 ) ``` **MODIS Terra/Aqua (250–500 m)** provides daily data throughout the flood period. NDWI or Modified NDWI can detect surface water quickly, ideal for tracking temporal evolution of the flood front. **Landsat 5/7 (30 m)** offers much better spatial resolution than MODIS, especially Band 5 (SWIR), which is sensitive to surface water. Landsat results can be overlaid with extracted channel networks to validate whether predicted channels correspond with actual flood paths. **ALOS-2 PALSAR / RADARSAT-2 (SAR):** SAR penetrates clouds, providing continuous data throughout the monsoon flood period with more accurate flood extent than optical sensors during rain. NASA Flood Observatory and Dartmouth Flood Observatory have already compiled SAR flood maps from this event. **JAXA GCOM-W AMSR2** provides daily flood inundation data. Though spatial resolution is relatively low (~10 km), it's useful for validating that the derived channel network has reasonable hydrological connectivity. ## Two-Level Validation Framework Design (Analogous to Kushiro) ```{r} #| echo: false # Create validation framework comparison validation_framework <- tibble( `Validation Level` = c( "Level 1: Pairwise Statistical Consensus", "Level 1: Pairwise Statistical Consensus", "Level 2: External Plausibility Validation", "Level 2: External Plausibility Validation", "Level 2: External Plausibility Validation" ), Metric = c( "Intersection over Union (IoU)", "F1-Score", "Channel-Flood Correspondence", "Flood Routing Plausibility", "Temporal Consistency Index" ), Formula = c( "IoU = TP / (TP + FP + FN)", "F1 = 2TP / (2TP + FP + FN)", "CFC = Channel pixels within flood extent / Total channel pixels", "FRP = Flood progression matches network flow direction", "TCI = Correlation between flow accumulation and flood timing" ), `Data Required` = c( "Multiple channel extractions from workflow combinations", "Same as IoU", "Extracted channel network + 2011 flood extent map", "Channel network + temporal flood progression (MODIS)", "Flow accumulation raster + multi-date flood maps" ), Purpose = c( "Measure agreement between different workflow outputs (internal consistency)", "Alternative metric for binary classification performance", "Check if predicted channels align with actual flood corridors", "Verify that flood propagation follows extracted network topology", "Validate that upstream areas flood before downstream areas" ) ) validation_framework %>% gt() %>% tab_header( title = "Two-Level Validation Framework", subtitle = "Analogous to Kushiro study structure, adapted for 2011 flood context" ) %>% tab_style( style = cell_text(weight = "bold"), locations = cells_column_labels() ) %>% tab_style( style = cell_fill(color = "#e8f4f8"), locations = cells_body( columns = `Validation Level`, rows = c(1, 2) ) ) %>% tab_style( style = cell_fill(color = "#fff3cd"), locations = cells_body( columns = `Validation Level`, rows = c(3, 4, 5) ) ) %>% cols_width( `Validation Level` ~ px(100), Metric ~ px(100), Formula ~ px(100), `Data Required` ~ px(100), Purpose ~ px(100) ) %>% tab_style( style = cell_text(size = px(10)), locations = cells_body() ) %>% tab_options( table.font.size = 11 ) ``` **Level 1:** Pairwise Statistical Consensus remains the same as before—comparing channel network outputs from different workflow combinations using the IoU metric from the Kushiro study: $$ IoU = \frac{𝑇P}{TP+FP+FN} $$ {#eq-1} **Level 2:** External Plausibility Validation uses 2011 flood inundation extent instead of Sentinel-2, based on the principle that **"a good channel network should predict flow paths consistent with actual flood extent."** This calculates spatial overlap between predicted channels and observed flood corridors: $$ ChannelFlood Correspondence = \frac{Channel Pixels Within Flood Extent}{Total Channel Pixels} $$ A new metric appropriate for the 2011 flood context is added: Flood Routing Plausibility, which checks whether the extracted network can explain the propagation of the flood front from upstream to downstream. ```{r} #| fig-height: 7 #| include: false # Visualize validation framework concept set.seed(123) # Simulate data for visualization n_workflows <- 12 workflow_names <- paste0("W", 1:n_workflows) # Level 1 metrics (pairwise IoU) iou_matrix <- matrix(runif(n_workflows * n_workflows, 0.4, 0.9), nrow = n_workflows) diag(iou_matrix) <- 1 # Level 2 metrics level2_metrics <- tibble( Workflow = workflow_names, Channel_Flood_Correspondence = runif(n_workflows, 0.5, 0.95), Flood_Routing_Plausibility = runif(n_workflows, 0.45, 0.90), Temporal_Consistency = runif(n_workflows, 0.40, 0.85) ) # Create heatmap for Level 1 iou_long <- as.data.frame(iou_matrix) %>% mutate(Workflow1 = workflow_names) %>% pivot_longer(cols = -Workflow1, names_to = "Workflow2", values_to = "IoU") %>% mutate(Workflow2 = rep(workflow_names, each = n_workflows)) p1 <- ggplot(iou_long, aes(x = Workflow1, y = Workflow2, fill = IoU)) + geom_tile(color = "white") + scale_fill_gradient2(low = "#d73027", mid = "#fee090", high = "#1a9850", midpoint = 0.7, limits = c(0.4, 1)) + labs(title = "Level 1: Pairwise Statistical Consensus (IoU)", x = NULL, y = NULL) + theme_minimal() + theme(axis.text.x = element_text(angle = 45, hjust = 1), plot.title = element_text(face = "bold", size = 11)) # Create bar chart for Level 2 level2_long <- level2_metrics %>% pivot_longer(cols = -Workflow, names_to = "Metric", values_to = "Score") p2 <- ggplot(level2_long, aes(x = Workflow, y = Score, fill = Metric)) + geom_col(position = "dodge", alpha = 0.8) + scale_fill_brewer(palette = "Set2") + labs(title = "Level 2: External Plausibility Validation", subtitle = "Using 2011 Flood Inundation Data", x = "Workflow Configuration", y = "Validation Score") + theme_minimal() + theme(axis.text.x = element_text(angle = 45, hjust = 1), plot.title = element_text(face = "bold", size = 11), legend.position = "top") # Combine plots p1 / p2 + plot_annotation( title = "Two-Level Validation Framework for Chao Phraya", subtitle = "Adapting Kushiro methodology to 2011 flood context", theme = theme(plot.title = element_text(size = 14, face = "bold")) ) ``` ## Clear Research Novelty and Contribution The Kushiro study's novelty lies in its workflow evaluation framework for UAV-LiDAR in wetland areas. To publish Chao Phraya work, distinct novelty is needed: ```{r} #| echo: false # Create novelty comparison table novelty_table <- tibble( `Novelty Aspect` = c( "1. Framework Transferability across Data Availability Contexts", "2. Flood Inundation as Hydrological Validation", "3. Human-Modified vs. Natural Channel Networks" ), `Kushiro Baseline` = c( "Dual-validation framework applied to high-resolution LiDAR in natural wetland", "Contemporary Sentinel-2 imagery for visual validation of channel presence", "Natural channel network (1st-3rd order streams) in pristine wetland" ), `Chao Phraya Innovation` = c( "Same framework applied to multi-source DEMs in data-poor context\nDEM source diversity replaces ground filtering diversity\nDemonstrates methodological robustness across data types", "Historical flood extent (2011) as proxy for network plausibility\nTemporal flood progression validates network topology\nSystematic use of flood data rarely done in channel extraction", "Extensive human-created canal network (khlongs)\nComplex mix of natural channels and irrigation infrastructure\nEvaluate automated extraction capability for both types" ), `Scientific Contribution` = c( "Methodological framework proves applicable beyond original context\nProvides template for other data-limited regions\nValidates dual-validation concept as generalizable", "Establishes flood inundation as valid hydrological proxy\nAddresses temporal gap between DEM and validation data\nCreates methodology for post-event validation", "Addresses understudied problem of mixed natural-anthropogenic networks\nRelevant for heavily modified river basins globally\nQuantifies extraction performance by channel type" ) ) novelty_table %>% gt() %>% tab_header( title = "Research Novelty and Contribution", subtitle = "Distinct contributions beyond Kushiro framework" ) %>% tab_style( style = cell_text(weight = "bold"), locations = cells_body(columns = `Novelty Aspect`) ) %>% tab_style( style = cell_fill(color = "#e8f4f8"), locations = cells_body(columns = `Novelty Aspect`) ) %>% cols_width( `Novelty Aspect` ~ px(150), `Kushiro Baseline` ~ px(150), `Chao Phraya Innovation` ~ px(150), `Scientific Contribution` ~ px(150) ) %>% tab_style( style = cell_text(size = px(10)), locations = cells_body() ) %>% tab_options( table.font.size = 11 ) ``` **Novelty 1:** Framework Transferability across Data Availability Contexts demonstrates that the dual-validation framework from Kushiro can transfer to more data-poor situations, where DEM source diversity replaces ground filtering diversity and flood inundation maps replace contemporaneous satellite imagery. **Novelty 2:** Flood Inundation as Hydrological Validation uses historical flood extent as a proxy for network plausibility—something few studies have done systematically. The challenge is explaining the temporal gap between DEM acquisition and the flood event (DEM may be from a different year than 2011). **Novelty 3:** Human-Modified vs. Natural Channel Networks: The Chao Phraya has an extensive human-created canal network, making flow routing far more complex than Kushiro. Evaluating how well automated extraction can recover both natural channels and anthropogenic canals is scientifically valuable. ## Cautions and Limitations ```{r} #| echo: false # Create limitations table limitations <- tibble( `Limitation` = c( "Vertical Resolution Mismatch", "Temporal Inconsistency", "Scale Dependency" ), `Problem Description` = c( "In Chao Phraya floodplain, elevation variation between channels and floodplains may be only 0.5–2 m\nThis falls within SRTM's RMSE range (3-5 m)\nCreates high uncertainty in channel delineation from DEM", "DEM captured before 2011 (SRTM: 2000, ALOS: 2006-2011)\nFlood event occurred in 2011\nCanals may have been dug or rice fields filled during interval\nLandscape changes not reflected in DEM", "Flow accumulation threshold of 5,000 m² used in Kushiro at 5 m resolution cannot be directly applied\n30 m resolution in Chao Phraya requires different thresholds\nDrainage area to classify must be recalculated" ), `Impact on Results` = c( "Channel detection uncertainty increases dramatically\nFalse positives and false negatives both likely\nExtracted network may miss shallow channels\nMay incorrectly identify non-channel features", "Validation may show poor correspondence in areas of landscape change\nExtracted network may not match 2011 flood paths if canals changed\nReduces reliability of flood-based validation", "Inappropriate thresholds lead to over/under-segmentation\nNetwork density may not match reality\nComparisons with Kushiro results require careful normalization" ), `Mitigation Strategy` = c( "Use multi-source DEM consensus to reduce individual source bias\nPerform comprehensive sensitivity analysis\nQuantify uncertainty ranges for all results\nClearly report confidence levels by terrain type", "Document all known landscape changes between DEM date and 2011\nUse multiple temporal DEM sources if available\nFocus validation on stable landscape areas\nExplicitly acknowledge temporal mismatch in limitations", "Calculate new thresholds based on drainage area analysis\nPerform scale-dependent validation\nTest multiple threshold values systematically\nReport results as functions of scale parameters" ) ) limitations %>% gt() %>% tab_header( title = "Cautions and Limitations", subtitle = "Critical challenges requiring careful handling" ) %>% tab_style( style = cell_text(weight = "bold"), locations = cells_body(columns = Limitation) ) %>% tab_style( style = cell_fill(color = "#f8d7da"), locations = cells_body(columns = Limitation) ) %>% cols_width( Limitation ~ px(150), `Problem Description` ~ px(150), `Impact on Results` ~ px(150), `Mitigation Strategy` ~ px(150) ) %>% tab_style( style = cell_text(size = px(10)), locations = cells_body() ) %>% tab_options( table.font.size = 11 ) ``` The biggest problem is vertical resolution mismatch: in the Chao Phraya floodplain, elevation variation between channels and floodplains may be only 0.5–2 m, which falls within SRTM's RMSE range. This creates high uncertainty in channel delineation from DEM, requiring clear sensitivity analysis and uncertainty quantification. Another issue is temporal inconsistency between the DEM (captured before 2011) and the flood event (2011), especially in areas where canals were dug or rice fields filled during that interval. This must be considered and documented in the limitations. Finally, scale dependency is critical because the flow accumulation threshold of 5,000 m² used in Kushiro at 5 m resolution cannot be directly applied to 30 m resolution in Chao Phraya. New thresholds must be calculated based on the drainage area to be classified. ```{r} #| fig-height: 5 #| include: false # Visualize scale dependency problem threshold_data <- tibble( Resolution = c(5, 10, 30, 90), Min_Detectable_Area_m2 = c(25, 100, 900, 8100), Kushiro_Threshold_m2 = 5000, Equivalent_Cells = Kushiro_Threshold_m2 / (Resolution^2) ) %>% mutate( Feasible = Equivalent_Cells >= 4, # At least 4 cells needed Label = paste0(round(Equivalent_Cells, 1), " cells") ) ggplot(threshold_data, aes(x = factor(Resolution), y = Equivalent_Cells, fill = Feasible)) + geom_col(alpha = 0.8) + geom_text(aes(label = Label), vjust = -0.5, fontface = "bold") + geom_hline(yintercept = 4, linetype = "dashed", color = "red") + annotate("text", x = 4.5, y = 4, label = "Minimum 4 cells", color = "red", hjust = 1, vjust = -0.5) + scale_fill_manual(values = c("TRUE" = "#1a9850", "FALSE" = "#d73027"), labels = c("Too few cells", "Sufficient cells")) + labs( title = "Scale Dependency Problem: Kushiro Threshold Applied to Different Resolutions", subtitle = "5,000 m² drainage area threshold from Kushiro study", x = "DEM Resolution (m)", y = "Number of Cells Representing 5,000 m²", fill = "Feasibility", caption = "At 30 m resolution (Chao Phraya), 5,000 m² = only 5.6 cells\nNew thresholds must be calculated based on actual drainage area characteristics" ) + theme_minimal(base_size = 12) + theme( legend.position = "top", plot.title = element_text(face = "bold", size = 13), plot.caption = element_text(hjust = 0, face = "italic", size = 9) ) ``` ## Integration with RRI Hydrological Model ### Complete Workflow Diagram ```{r} #| fig-height: 12 #| fig-cap: "Complete workflow from data inputs through validation to final output" grViz(" digraph workflow { graph [rankdir=TB, fontname='Arial', bgcolor='#f8f9fa', splines=ortho, nodesep=0.8, ranksep=1.2] node [shape=box, style='rounded,filled', fontname='Arial', fontsize=11, margin=0.2] edge [fontname='Arial', fontsize=10, penwidth=2] // Input data nodes gpm [label='GPM IMERG\\nFinal Run\\n(0.1° × 30min)', fillcolor='#e3f2fd', color='#1976d2', penwidth=2] srtm [label='SRTM v3\\n(30 m)', fillcolor='#fff3cd'] alos [label='ALOS AW3D30\\n(30 m)', fillcolor='#fff3cd'] cop [label='Copernicus DEM\\n(30 m)', fillcolor='#d4edda', color='#28a745', penwidth=2] tandem [label='TanDEM-X\\n(12 m)', fillcolor='#fff3cd'] // DEM processing dem_merge [label='Multi-source DEM\\nEnsemble', fillcolor='#ffe0b2', shape=diamond] // Kushiro Framework kushiro [label='Workflow Evaluation\\n(Kushiro Framework)\\n\\n• DEM Source: 4 options\\n• Sink Fill: 2 methods\\n• Flow Dir: 2 methods\\n• Flow Accum: 3 thresholds\\n\\n= 48 combinations', fillcolor='#f3e5f5', color='#7b1fa2', penwidth=2] // Channel extraction channel [label='Channel Network\\nExtraction\\n(Optimal Workflow)', fillcolor='#c8e6c9'] // Validation Level 1 validation1 [label='Level 1 Validation\\nPairwise IoU\\n(Internal Consistency)', fillcolor='#ffccbc', shape=parallelogram] // RRI setup rri_setup [label='RRI Model Setup\\n\\n• Grid: 250 m\\n• Channel: Extracted network\\n• Parameters: Literature\\n• Rainfall: GPM IMERG', fillcolor='#b2dfdb', color='#00695c', penwidth=2] // RRI simulation rri_run [label='RRI Simulation\\n2011 Flood Event\\nJul-Dec 2011', fillcolor='#b2ebf2'] // Output sim_flood [label='Simulated Flood\\nInundation Map\\n(Daily extent)', fillcolor='#c5cae9'] // Validation data obs_flood [label='Observed 2011\\nFlood Maps\\n(MODIS/Landsat/SAR)', fillcolor='#f8bbd0', color='#c2185b', penwidth=2] // Validation Level 2 validation2 [label='Level 2 Validation\\n(External Plausibility)\\n\\n• Flood extent accuracy\\n• Temporal consistency\\n• Routing plausibility', fillcolor='#ffccbc', color='#d84315', penwidth=2, shape=parallelogram] // Final output final [label='Validated Flood\\nInundation Map\\n+ Uncertainty', fillcolor='#a5d6a7', color='#2e7d32', penwidth=3, style='rounded,filled,bold'] // Connections - Input stage srtm -> dem_merge alos -> dem_merge cop -> dem_merge tandem -> dem_merge // DEM to Kushiro dem_merge -> kushiro // Kushiro to Channel kushiro -> validation1 [label='48 networks'] kushiro -> channel [label='Best workflow', color='#2e7d32', penwidth=2.5] validation1 -> channel [style=dashed, label='Quality check'] // Channel to RRI channel -> rri_setup [label='Drainage network'] gpm -> rri_setup [label='Precipitation\\nforcing'] // RRI flow rri_setup -> rri_run rri_run -> sim_flood // Validation sim_flood -> validation2 obs_flood -> validation2 [label='Ground truth'] // Final output validation2 -> final [label='Calibrated', color='#2e7d32', penwidth=2.5] // Subgraph grouping subgraph cluster_input { label='Input Data' style=dashed color='#616161' fontsize=12 gpm srtm alos cop tandem } subgraph cluster_topo { label='Topographic Processing' style=dashed color='#7b1fa2' fontsize=12 dem_merge kushiro validation1 channel } subgraph cluster_hydro { label='Hydrological Modeling' style=dashed color='#00695c' fontsize=12 rri_setup rri_run sim_flood } subgraph cluster_validation { label='Validation & Output' style=dashed color='#d84315' fontsize=12 obs_flood validation2 final } } ") ``` ## RRI Model Configuration The Rainfall-Runoff-Inundation (RRI) model is a 2D physically-based distributed hydrological model that simultaneously solves: - Rainfall-runoff on hillslope grid cells - 1D river flow in channel network - 2D overland flow on floodplains - Lateral exchange between river and floodplain ```{r} rri_setup_params <- tibble( Parameter = c( "Model Domain", "Grid Resolution", "Simulation Period", "Time Step", "Channel Network", "Rainfall Input", "Boundary Conditions", "Initial Conditions", "Output Frequency" ), Value = c( "Chao Phraya Basin", "250 m × 250 m", "July 1 - December 31, 2011 (184 days)", "60 seconds (adaptive)", "From optimal Kushiro workflow", "GPM IMERG Final Run (30-min)", "C.2 station discharge (upstream)", "Dry condition (water depth = 0)", "Daily inundation maps + hourly hydrographs" ), Justification = c( "Covers 2011 flood extent (~20,000 km²)", "Balance between accuracy and computational cost", "Captures entire 2011 monsoon flood event", "Ensures numerical stability (CFL condition)", "Validated drainage network from framework", "Best available gridded precipitation data", "Chiang Mai station provides upstream inflow", "Monsoon onset marks simulation start", "Daily for validation, hourly for dynamics" ) ) create_table( rri_setup_params, title = "RRI Model Configuration", subtitle = "Setup for 2011 Chao Phraya flood simulation" ) %>% {if(knitr::is_latex_output()) { column_spec(., 1, bold = TRUE, width = "3.5cm") %>% column_spec(., 2, width = "4cm") %>% column_spec(., 3, width = "4cm") } else { tab_style(., style = cell_fill(color = "#e3f2fd"), locations = cells_body(columns = Parameter) ) %>% tab_style(., style = cell_text(weight = "bold"), locations = cells_body(columns = Parameter) ) %>% cols_width( Parameter ~ px(180), Value ~ px(260), Justification ~ px(280) ) }} ```