Exploring the Feasibility of Mitigating Flood Hazards by an Existing Pond System in Taoyuan, Taiwan

Abstract

Changes in the global climate have induced densified rainfall and caused natural hazards across the world in recent years. Formed by a central mountain range and a corridor of alluvial plains to the west, Taiwan is at risk of flood hazards owing to its low-lying lands as well as the distinct seasonality of rainfall patterns. The rapid discharge of surface runoff and a growing number of impervious surfaces have also increased flood hazards during recent typhoon landfalls. A century ago, ancestors in Taoyuan City constructed a system of water channels composed of thousands of ponds to fulfill the needs of agriculture and aquaculture. During the expansion of urban areas, land reformation replaced a majority of earlier ponds with residential and industrial zones. However, the remaining ponds could potentially serve as on-site water detention facilities under the increasing risk of floods. In this research, we first renewed an outdated pond database by deploying a novel unmanned aerial vehicle (UAV) system with a micro-sonar to map the bathymetry of 80 ponds. Next, a simplified inundation model (SPM) was used to simulate the flood extent caused by different scenarios of rainfall in Bade District of Taoyuan City. Assuming that extremely that heavy rainfalls at 25, 50, 75, and 100 mm occurred in a very short period, the flood area would decrease by 96%, 75%, 52%, and 37%, respectively, when the ponds were preparatorily emptied.

1. Introduction

A network of ponds scattered over a 30 $\times$ 30 km area is an iconic landscape in northern Taiwan (Figure 1). The ancient Shihmen river, flowing through the center of current Taoyuan City and forming the alluvial plain, was captured by the northbound river and reduced water levels at lower reaches more than 30,000 years ago [1]. The ancestors who arrived here centuries ago had settled along the river bank or areas with accessible groundwater. Along with the growing population, residents who in the early stage farmed with precipitation and natural watercourses had to build water facilities in the last century [2]. Freshwater supply had become an issue, so the ancestors irrigated with ponds and ditches. Although Taoyuan’s geographical environment is unsuitable for dams as the average slope of streams is 1/40–1/120 [2], the soil type composed of laterite and loess is conducive to constructing artificial water storage facilities. As time went on, some of the ponds functioned as small reservoirs in the water source management system and were linked with rivers and streams. These ponds’ functionalities have become an intricate system for irrigation, drainage, wetland conservation, and aquafarming [3] (Figure 1, inset figures).

1.1. Pond Network in Taoyuan

Recently, the demands for agricultural production decreased, and the rice fields in many irrigated areas were progressively replaced by urban areas [2,4], due primarily to the rapid development of industry and commerce. From 2002 to 2018, the area of cultivated land decreased from 39,608 to 31,896 ha in Taoyuan [5]. According to the latest mapping in 2011, there are still 2851 existing ponds, as delineated by the Department of Water Resources, Taoyuan (TYWR) from aerial photography. Because of the cutback of water demand for agriculture, some channel sectors within the network gradually disintegrated where ponds were abandoned. These diminishing ponds have limited storage capacity due to sedimentation. Wang (2013) [6] suggested that converting fish ponds into flood detention basins could mitigate flood hazards regardless of whether the accumulated precipitation is 150 mm or as high as 1200 mm. Wang & Chang (2016) [7] also explored the spatial connection between the ponds in Taoyuan and investigated the possibility of flood detention. They assumed an average depth of 3.4 m for 497 ponds and concluded that using this existing configuration could effectively reduce flood risks. For a daily accumulated rainfall of 200 mm, the ponds in Dayuan District, near Taiwan Taoyuan International Airport in Figure 1, could reduce 55% of the water from a simulated flood event [7].

Unfortunately, most ponds have very limited information, even in the government database. Incomplete background knowledge is a major drawback for an overall assessment and planning. Therefore, to fully explore the potential usage of their capacities, detailed geometric parameters are needed. This study, therefore, first intended to develop a complete digital elevation model (DEM) to seamlessly connect the pond floor and the neighboring land for later use in flood simulation. As demonstrated in Bade District of Taoyuan City in Figure 2a, the DEM from the Ministry of the Interior (Taiwan) is unable to reveal detailed pond shapes. As the close-up in the orange box in Figure 2a demonstrates, the water surface is hydro-flattened during the editing of airborne LiDAR data, and thus the elevation is similar to that of the neighboring land (Figure 2b). Therefore, as demonstrated in Figure 2c, the DEM used in the following simulation was modified according to either a single depth value in the database or the bathymetric maps updated by our fieldwork.

1.2. Methods for Measuring Inland Waterbodies

To measure point depth over a lake, surveyors used to deploy sounding rods, sounding leads and other tools lowered from a vessel. Lately, optical and acoustic sounding sensors were developed by counting the elapsed time of wave propagation. There are active and passive mechanisms in optical sounding [8]. The most popular device for shallow water areas is the airborne laser bathymetry (ALB) system, which estimates the range based on the bidirectional propagation time of short pulses between the surface and bottom. ALB can nominally survey depths to 60 m depending on the water clarity [9]. The laser is split by an optical coupler into an infrared beam and a green beam. The infrared beam is used to detect water surfaces because of its poor penetration, while the green beam zig-zagging across-track aims to measure depths. The advantage of active ALB is the insensitivity of sunlight angles and the reflections of wave surfaces as compared with passive optical sounding, as well as its workability under insufficient light sources [10,11]. Hilldale and Raff (2008) [12] collected 220 km of ALB data at the Yakima Basin in Washington and the Trinity River Basin in Texas to perform surveys of two riverbeds. Comparing LiDAR data and the single-beam echosounder with the real time kinematic technique of the global positioning system (RTK-GPS), the results showed that the standard deviation is at 0.3–0.42 m level when the slope is less than 20%, and it may be up to 0.63 m otherwise.

In contrast, passive optical sensor onboard satellites can estimate water column thickness from the light attenuation between spectral bands. This approach reduces costs in the field and avoids the inaccessibility of remote areas. Stumpf et al. (2003) [13] proposed a log-ratio transformation utilizing spectral attenuation based on the Beer–Lambert Law. By comparing IKONOS satellite images with depth measurements, a linear variation algorithm and an empirical proportional variation algorithm are constructed. An alternative way to exploit optical images is based on the hypsometric approach. Getirana et al. (2018) [14] used Landsat historical images and SARAL/AltiKa altimetry satellite to build a slope elevation model for Lake Mead, USA. They further used flow direction and neighboring hydrological formations to derive full reservoir bathymetry. In their results, interpolated DEM in consideration of the upstream and downstream elevation agreed well with surveys from scan-sonar and chirp seismic reflection.

Besides, acoustic methods had also been widely used to detect and locate objects/obstacles in the water. Sound navigation and ranging (Sonar) can be divided into active and passive types based on function and mode. Sonar systems are categorized as, for example, side-scan sonar, multibeam sonar, acoustic communication system, positioning system, acoustic Doppler system, and acoustic tomography network. Active sonar is mostly operated for bathymetric mapping, fish detection, and sediment profiling [15]. However, these systems are primarily used in waterbodies where the cruising of vessels allows. For a small pond or a narrow channel, the mobility of ships is constrained and the cost of a survey campaign may not be economically reasonable.

1.3. A Novel Bathymetry Technique

With the boom in commercial-grade drones, various ingenious applications have been developed thanks to their outstanding maneuverability in the field. The unmanned aerial vehicle (UAV) has been used to broaden research areas in agriculture, forestry, flood monitoring, and geohazard assessment [16,17]. The DEM generated from aerial photos using the structure from motion (SFM) technique has a high degree of conformity with aerial LiDAR data [18]. Javemick et al. (2014) [19] used stereo-paired images taken at 600 m and 800 m above the ground and performed SFM to construct a DEM. The geolocation errors were 0.04 m in planar, and 0.10 m in vertical directions. UAVs have also been utilized for bathymetric mapping, outperforming traditional approaches in efficiency and accessibility. Alvarez et al. (2018) [20] conducted shallow water depth mapping in a small reservoir in Oklahoma, USA, where the study area was about 28,000 m². An echosounder mounted onboard a floating platform was towed by a UAV to measure water depth. They retrieved scattered point depths and applied cluster analysis to distinguish areas with large standard deviations. The underwater topography map was obtained by spatial fitting of the measured depth points, and a standard deviation of 0.37 m was reached between the fitted terrain and reference data. Bandini et al. (2018) [21] combined a UAV, a sonar, an inertial measurement unit (IMU), a SONY RX-100 camera, and a GNSS module to correct the distortion and tilt. Similarly, the entire sonar system was towed by a UAV and the result showed an accuracy of ~2.1% of actual depth with a maximum depth of 35 m. Many innovative approaches developed in the abovementioned experiments considerably reduce costs compared with traditional methods.

Due to the inaccessibility of several unmanaged ponds in Taoyuan, we also needed a mobile platform with good flexibility when hovering above a pond. Therefore, a novel measuring system was developed to map the depth of ponds. This research first renewed the outdated depth of ponds with a mobile sounding system comprising a UAV, a smartphone, and a micro-sonar. The innovation of this UAV module was to hang the sonar by an aerial vehicle to skim the surface of the water, without the floating platform that is usually restricted by launching and mobility in a small pond of a few thousand square meters. Also, this is the first time that a group of pond models was integrated with an adjacent DEM to simulate the overland flow under different rainfall scenarios using the simplified inundation model (SPM) [22].

2. Methodology

2.1. Workflow

The workflow of this study is outlined in Figure 3. The first step was to select the study area and ponds to be measured. Here, we picked 80 ponds in Taoyuan City, of which 15 ponds are located in Bade District. To update the geometry of each site, four main tasks were conducted, including the deployment of control points, operation of aerial photography, UAV bathymetry, and other ground surveys. To model the neighboring land, aerial photos taken within a 200 m buffer were stereo-paired to build a DEM constrained by GNSS control points with precise point positioning (PPP) solutions. To model the underwater part, the measured depth points were used to interpolate a mesh grid of bathymetry in consideration of slope and channel geometry. For the land area outside the buffer or for the pond without fieldwork, the 20 m digital terrain model (DTM) released by the Ministry of the Interior (MOI), Taiwan (https://dtm.moi.gov.tw (accessed on 1 July 2022), version 2018) was used. The DTM was produced by a nationwide airborne LiDAR campaign, and the specified accuracy of LiDAR measurement was 0.25 m in vertical for each 1 m grid comprising more than 2 LiDAR points. The joint terrain model, composed of an underwater DEM, a land DEM, and the MOI DEM, was assembled and bilinearly resampled into a 0.5 m resolution. The next step was to use this integrated DEM for SPM flood simulation. Finally, the floodwater was assumed to be dissipated into the ponds via a virtual network, and the contribution of the ponds to flood detention was assessed.

2.2. Fieldwork Procedure

2.2.1. Modeling of Terrain DEM

For fieldwork in the selected pond, one operator controlled a DJI Phantom 4 Pro (DJI-P4P) UAV to take aerial photos for reconstructing DEM over the land area. A flight height of 60 m and an overlap rate of 70% were set with a ~200 m buffer area outside the pond. To ensure the quality of geolocation for the orthorectified images, 3–5 ground control points were positioned by a u-blox C94-M8P GNSS receiver for more than 30 min. This GNSS chip provides single-frequency phase pseudo-ranges compatible with GPS, GLONASS, and Galileo. GNSS observation files are imported to RTKLIB [23] open-source software to solve Precise Point Positioning (PPP) coordinates. With PPP, the accuracy could be increased to a decimeter level for phase pseudo-ranges, owing to the reduction of biases and clock/orbital errors.

2.2.2. Modeling of Pond DEM

A DJI Phantom 3 Advanced (DJI-P3A) UAV was deployed to carry a bathymetric module along the planned route above the water surface. Under moderate environmental conditions, including good GPS visibility, calm wind speed, and low RF interference, the UAV can stay within 1.5 m in horizontal and 0.1 m in vertical directions. The maximum load under DJI-P3A is limited to ~500 g based on previous tests [24]. Hence, a lightweight (89 g) Ling-Hui micro-sonar and a smartphone can be hung by the UAV with a fishing string 1.5 m long (Figure 4). The ranging capability of this leisure-grade fish finder is between 0.6 m and 30 m and the working temperature is <40 °C. Although the main application of this low-cost (~70 USD) micro-sonar is for fish detection, it is possible to estimate the depth owing to its sensitivity at ~10 cm level [24]. The sonar data were transmitted to the smartphone through Wi-Fi, with a maximum transmission distance of 50 m.

The operator first planned a grid of points on the pond where the UAV could stay for 1 min. Water depths displayed in the mobile application (App) that came with the sonar were screenshotted every 2 s, with a total of 25–30 repetitions for each point. After removing the first and last 15 s when sonar readings were not yet stabilized, approximately 10–12 redundant measurements were retained for computing water depth. As this low-cost sonar does not have a positioning module, the smartphone-recorded National Marine Electronics Association (NMEA) file was used for geolocations. The assisted global positioning system (A-GPS), with good coverage of cellular networks, outperforms the code positioning by GPS [25]. This approach also avoids positioning errors due to the obscured sky visibility and ionospheric effects. The images screenshotted in the mobile App contains the resolved depth information. Here, we used an automatic recognition module coded in MATLAB to digitize water depth and timestamps recorded in each image. To produce DEM over the land area, a Pix4D Mapper was adopted to process aerial photos and control points to establish DEM and a mosaicked image. The water area mask was visually delineated from the mosaic image.

Among several interpolation strategies for modeling terrain, the inverse weighted distance (IDW) and the ordinary Kriging (OK) are most commonly favored to estimate continuous surfaces [26]. A study in Saldanha Bay further suggested that IDW consistently performed better than OK across multiple interpolation tests [27]. Based on the assumption that the bottom sedimentation of the ponds is smooth and the depth points are evenly distributed, IDW is reasonable in the presentation of floor topography. Hence, to fit the surface of underwater terrain from pointwise measurements, we adopted IDW as the spatial interpolation method. The formula of IDW [28,29] is

$$D\left(i,j\right)=\frac{{{\displaystyle \sum }}_{x=1}^n \frac{D_x}{d{\left(i,j\right)}_x}}{{{\displaystyle \sum }}_{x=1}^n \frac{1}{d{\left(i,j\right)}_x}}=\frac{{{\displaystyle \sum }}_{x=1}^n{D}_{x}W{\left(i,j\right)}_x}{{{\displaystyle \sum }}_{x=1}^{n}W{\left(i,j\right)}_x}$$

(1)

where $D\left(i,j\right)$ is the depth of an unknown point, $D_x$ is the value of a measured point and $d{\left(i,j\right)}_x$ is the Euclidean distance from the unknown point $D\left(i,j\right)$ to the x^th measured point. The weight in IDW is thus $W_i$, such as 1 divided by $d{\left(i,j\right)}_x$. IDW searches adjacent data at the pixel to be interpolated, and calculates the weighted average inversely proportional to the distance. Following that, the slope of the enclosing dyke is considered in generating an underwater model. We assumed that the dyke had a uniform slope as they were mostly renovated by a precast concrete form (Figure 5). The combination of slope measurement and IDW can present a completely seamless model rather than just using one of them. Finally, the joint pond model can be obtained by combining the underwater model with the land model.

To calculate the maximum storage (V_max) of each site, dyke or water gate height measured in situ was considered as the maximum water level. Equation (2) calculates water storage for each 0.5 $\times$ 0.5 m pixel (row i from 1 to n and column j from 1 to m) falling within the pond mask:

$${\mathrm{V}}_{\max }={{\displaystyle \sum }}_{j=1}^n{{\displaystyle \sum }}_{i=1}^{m}A\left(i,j\right)\times \left[D\left(i,j\right)+O_k\right]\times M\left(i,j\right)$$

(2)

where A is the unit area of each pixel (0.25 m²), D is the depth, $O_k$ $\left(k\in \left\{1,2\right\}\right)$ is the extra height measured in Figure 5, which is the lower one of height to the top of concrete dyke ($O_1$) or the slab height of the water gate ($O_2$). M is the pond mask, set to 1 within the pond and 0 for the land area. Figure 6 displays two examples of pond model in Yangmei District (YM145) and Bade District (BD033).

2.3. Flood Simulation

2.3.1. Simplified Inundation Model (SPM)

SPM [23] refers to a set of simple physical properties that were developed to simulate flooding extent by using DEM as the sole input parameter. The model requires rainfall and a threshold that represents the acceptable difference between total flood and simulated flood. It is suggested to set 10% of the rainfall remaining on the ground to calculate the extent and depth of a flooding event [22]. Compared to other simulation methods, SPM produces results promptly. The flow direction is first calculated and the rows of cell arrays are sequenced, and cells located at the end of arrays would be considered a high contribution. The flood would be drained to a lower elevation or a higher contribution, and the amount of water is in line with the preset rainfall and threshold. D-infinity [30] is used to transport water from flooding cells to surrounding cells. When the flow direction is between two cells, water is distributed according to the slope as weight because the downstream areas also have a higher risk of flood. As demonstrated in Figure 7a, a model with nine cells represents the main terrain formation. As the water level increases from cell #5 in panel (b), eight possible flowing directions are displayed as orange arrows while the steepest slope is marked as the red arrow. Angles a and b between two flowing directions to cell #1 and #2 are used as the weight to allocate water from the center cell, in terms of b/(a + b) and a/(a + b), respectively (Figure 7b). This process will continue until the water level is balanced among surrounding cells (Figure 7c) [22,30].

The convenience and outstanding speed of SPM have been demonstrated in the literature [31], highlighting that SPM using only elevation data for simulation was superior in areas without detailed hydrological parameters. SPM is much faster than other models in both data collection and processing efficiency, and the precision when performing on the town- and village-scale could reach an acceptable level, where the fit indicator (the overlapped flooded area divided by the total flooded area) is greater than 0.69 [31].

2.3.2. A Virtual Channel Scheme to Dissipate Floodwater

After simulation using SPM, the initial flood zones and their depths were estimated. However, if we assumed a virtual channel network existed between the flood-prone areas and the neighboring ponds, the flooded area might be further reduced based on a redistribution scheme. First, we assumed that the pond collocated with a flooded area would be filled. Next, the disconnected flooded areas in the SPM results were divided into patches. Starting from the lowest elevation, the flooded patches, one by one, the ponds not full and located downslope were searched. The flooded patch was first drained to the candidate with the shortest Euclidean distance. When the pond was full, the flooded patch found the next closest one to share floodwater, until no more successors located downslope had room to accommodate them. It should be understood that although the virtual channels do not exist in the field at this moment, based on our simulation, the demand for a drainage system could be visualized and the route could be considered for future plans.

3. Result

3.1. Validation of UAV Depth Measurements

A ship survey campaign was conducted in a pond #GI215 in Guanyin District Taoyuan. The Otter Unmanned Surface Vehicle (USV) carrying a Norbit iWBMS multibeam echosounder was deployed to scan the bathymetry. As shown in Figure 8a, the detailed elevation from the top of the dyke is revealed in USV sidescan results, with a range resolution of <1 cm. To compare the depth with measurements from our UAV module, 14 points skimmed by UAV are marked as black circles in Figure 8a,b, while the bathymetry in Figure 8b was modeled by the IDW strategy. The accuracy of UAV data, in terms of standard deviation and correlation coefficient as compared against USV data over those 14 depth points, achieved 0.12 m and 0.86, respectively. The error is about 4.8% of depth if we took 2.5 m of the averaged depth as a benchmark.

3.2. Validation of Integrated Pond Models

In 2019, 80 ponds over the entire Taoyuan City were selected for fieldwork. Each location required 3–4 onsite staff to take 1–1.5 h of measurement. The detailed data processing and validation procedure can be found in [24,32]. The 80 models, as shown in Figure 9, have a range of water depth from 0.7 to 4.4 m and a range of storage from 2191 m³ to 155,564 m³. For comparing the configuration of ponds with historical data from TYWR, the correlation coefficient of water extent reached 0.99 (Figure 10a). However, in the comparison of water storage, two (CL143 and YM051, see #8 and #53 in Appendix A) had extraordinarily mismatched values. The depth calculated roughly from storage and the area in the TYWR record was just 0.07 m and 0.27 m for CL143 and YM051, respectively, which was far shallower than the common depth ranging between 1.2 m and 2.0 m. Therefore, the outliers were attributed to the significant underestimate of storage in the database. The correlation coefficient of storage without those two sites was improved from 0.70 to 0.83 (Figure 10b).

The results shown here have two-fold implications. First, the fidelity of our pond models should satisfy the needs of SPM simulation. Some depths in the database may be unreliable due to an inaccurate taping observation or a lack of updating recent dredging processes. The UAV sensing module introduced herein is flexible in cost, accuracy, and coverage. Second, the historical data need to be routinely updated to grasp the current status of bathymetry. Severe sedimentation is of great concern to some kinds of functionalities, for example, aquafarming and irrigation water storage.

3.3. Design of Drainage Channels from Virtual Network Dissipating Scheme

In this research, Bade District in Taoyuan City (Figure 1) was chosen for flood simulation because of its high level of urban development and the history of frequent flooding events. During short periods of heavy rainfall, floods often took place owing to an inadequate drainage system, even though the number and scale of ponds in Bade are larger than in other districts [33]. To assess the potential usage of the current infrastructure as a floodwater detention facility, SPM was applied in Bade with rainfall amounts of 25 mm, 50 mm, 75 mm, and 100 mm. Because no drainage conditions other than elevation are required in SPM, it has the great advantage of presenting short periods of heavy rainfall.

The initial flooded area and redistribution procedure of floodwater was exemplified by the 75 mm rainfall case. As the flood just happened, the area simulated by SPM was shown as blue patches in Figure 11a. Those submerged areas could be further dissipated through a virtual channel scheme defined in Section 2.3.2. Following the algorithm, the virtual network, or the lines connecting flooded patches to those ponds with unfilled space, is visualized as black lines in Figure 11a. It is noted that because some detention basins have been filled in the low-to-high elevation order, several flood patches need to be drained across the main slope direction (south to north). If the floodwater could be successfully redirected to the target ponds, the flooded areas could be reduced, as indicated by the red patches in Figure 11b. The flooded extent is much smaller than its initial form, since many ponds situated outside of flood-prone areas could potentially share stormwater under the assumption of appropriate drainage facilities. However, considering the actual constructional and operational difficulties of this virtual network, the black lines in panel (a) are simplified as three main channels, such as A to A’, B to B’, and C to C’ in panel (c), indicating three routes representative of the most demands in the simulation. It is observed that a cross-tributary design in the upstream (A-A’) and a new channel linking flooded areas to a huge pond (C-C’) or to a pond cluster (B-B’) are preferred for this purpose.

A similar simulation of the dissipating procedure was conducted for the other three rainfall scenarios and the results are illustrated in Figure 12. In this figure, the blue plus red patches indicate flooded areas initially simulated by SPM and the red patches are the ones left after virtual channel dissipating scheme, similar to an overlap of panel a and b in Figure 11. Another symbol introduced in this figure is the green patch, which is used to highlight the difference between using depth in the historical data (with green) and using the updated pond model (without green) in simulations. For example, in the 25 mm case, the green patches in the middle exist because old data underestimate pond storage, and thus the floodwater cannot be shared by the neighboring ponds. In contrast, by updating several ponds through our fieldwork (yellow filling) to the west, most sites were found bigger in their storage and hence could accommodate floodwater from those green patches. It is clear, especially in 25 mm and 50 mm rainfall scenarios. However, as rainfall increases, the water storage capacity is minor in the portion of floodwater, so the DEM correctness becomes less important. As a result, almost no green patches can be seen in 75 mm and 100 mm cases.

Because SPM gives the highest contribution to the lowest elevation in this region, the flood accumulates from the top (north) to the bottom (south) direction. When the rainfall amount is 25 mm, most of the flooded areas could be dissipated if there was a channel system linking ponds downslope of the flood hotspots. When rainfall amount reaches 50 mm, there could still be a significant reduction in the flooded areas. However, when rainfall amount reaches 75 mm or 100 mm, the contribution of ponds becomes progressively insignificant. Figure 13 shows the reduction percentage in flood areas and water volumes, based on four rainfall scenarios. The red solid/dashed line and blue solid/dashed line indicate the simulated flood area with/without our fieldwork and the simulated water volume with/without our fieldwork, respectively. When rainfall amounts increase from 25 mm to 50 mm, 75 mm, and 100 mm, reductions of flood area by the pond model updated in this study are 96%, 75%, 52%, and 37%, respectively. A similar trend is found in flood volume. When rainfall amounts increase in those four scenarios, reductions of flood volume by the updated pond model are 93%, 56%, 28%, and 14%, respectively. Overall, results from the updated pond model show up to 10% less flooded area and 3 to 7% less floodwater volume than the old model (TYWR) in four rainfall scenarios. Based on this simulation, the administrative agencies could have an overview of the first-order flood-prone areas in this district, and plan/renovate inter-pond channels that are close to the flooded patches. Meanwhile, those on-site tenants must carry out desilting work regularly, once the ponds are requisitioned to the flood detention network.

4. Discussion

This study utilized a novel water depth measuring module that could apply to small waterbodies difficult to be accessed or to be surveyed with traditional equipment. Meanwhile, the accuracy of UAV-assisted depth measurement was comparable with the reference data. The advantage of this approach is the capability of recovering underwater terrain of a pond insensible by optical remote sensing methods, or the ones full of aerators and wires unable to operate a ship. The method is particularly handy for quickly verifying the renovation of a pond. For example, the Xipo ecological park (labeled in Figure 11c, a combination of BD089 and BD090 in Appendix A) has 88,092 m³ of the estimated storage, which is close to the nominally new storage announced onsite (88,000 m³) [34] than the record in the outdated TYWR database (40,543 m³). It is also the main reason why several black lines in Figure 11a connect between flooded patches to the Xipo ecological park, such as the route C-C’ in Figure 11c. Actually, the three main routes suggested in Figure 11c are in accordance with a majority of built-up areas in Bade District, as shown in Figure A1 of Appendix B. It affirms the need for a new drainage system to reduce the risk of flood hazards over the densely populated area. In the future, when this procedure could be regularly operated, the tenant and owner (administrative agencies) should systematically monitor the timely change of pond sedimentation.

Moreover, the actual storage is crucial for the first crop of paddy rice cultivation (January–June) for irrigation purposes. Taiwan experienced a severe drought event in 2021 [35], which was the worst in 56 years, and forced almost all farmers to leave their land fallow in April and May. Therefore, preparatory storage of water in the pond system could mitigate the impact of drought events under any prompt water rationing schemes from the reservoir upstream. Based on Appendix A, the sum of estimated storage (2,928,301 m³) is greater than the TYWR record (2,647,510 m³) by 280,791 m³. If we took the 2500:1 as the ratio of water supply for rice cultivation [36], the storage estimation uncertainty is 112 tons of rice production in equivalent, which highlights the need for accurate quantification of capacity in this regard.

On the other hand, the pond system has an advantage in flood prevention. Although the terrain of Taoyuan city belongs to the tableland, for the time being, historical flood hazards had not caused a disastrous impact. Flood events in the record showed that regional flooding was mainly caused by insufficient drainage and large amounts of precipitation in a short period, which makes SPM applicable in this case. In this research, after updating the storage of 15 ponds in Bade District, as the ones starting with “BD” in Appendix A, the total water storage in Bade increases by 127,216 m³ (from 252,475 m³ to 379,691 m³), about 10% of the previous total volume. Our analysis shows that the virtual detention system could effectively reduce 83–93% and 45–56% of flood areas under 25 mm and 50 mm precipitation scenarios, respectively. However, the flooded area is much less reduced when the rainfall is 75 mm (28% reduction) and 100 mm (14% reduction) by using either new or old pond DEM. Another simulation (not shown here) indicates that if a total of 224 ponds larger than 400 m² in Bade could increase by 20% of its storage, the ratio of remaining flooded areas and volumes could be further decreased by 4% and 3.4%, respectively when rainfall is 100 mm.

Compared with various depth mapping methods at present, traditional methods such as surveying instruments are time-consuming and laborious or require the navigation of ships. Although the precision and resolution of LiDAR are more attractive, the instrument is expensive (updated every six years in Taiwan) and the operation is highly restricted by water quality. In this study, a UAV combined with sonar was proven to reduce time and labor costs [24]. Meanwhile, the UAV module shows high mobility characteristics that undoubtedly provide a novel method for water storage estimation. Although a few estimates in our research deviated from those in TYWR, attributed to seasonal and anthropogenic factors, the study showed that most results match well with the planned storage.

5. Conclusions

Flood hazard has become a deepening problem because of urban development and the loss of natural pervious surfaces. Asphalt in urban reduces seepage and allows heavy rainfall runoff to flood. Ancestors built thousands of ponds around current Taoyuan City to collect precipitation for irrigation. However, with the transformation from an agricultural to an industrial society, the watercourse has been deserted because of the lack of appropriate management and operation. The basic information of waterbodies, i.e., area and storage, has been poorly investigated and thus put huge uncertainty when using them for data analysis and simulation. In this research, a novel bathymetry technique, UAV with sonar and smartphone, was developed to perform bathymetry. Easily implementated with a few operators at each site, the fully integrated pond model with the neighboring land can be built for updating the outdated database. The system successfully detected two anomalies (CL143 and YM051) whose storages were apparently incorrect in the record. In the future, this system can be used to rapidly verify ponds whose single geometric parameter (area, depth, and volume) is not convertible by others.

In flood simulation, SPM showed the potential of a >70% reduction in flood areas in Bade when a sudden heavy rain of <50 mm occurred. The existing pond network in Bade may assist in retaining floodwater and mitigating the severity of flood hazards. However, it should be emphasized that the flood maps derived herein depend on the assumption of virtual channels between flood patches and ponds located downslope. The results could be a good reference for administrative agencies to optimize or reinvigorate abandoned channels/pipelines currently in the field.