skip to main content


Title: Accuracy-Constrained Efficiency Optimization and GPU Profiling of CNN Inference for Detecting Drainage Crossing Locations
The accurate and efficient determination of hydrologic connectivity has garnered significant attention from both academic and industrial sectors due to its critical implications for environmental management. While recent studies have leveraged the spatial characteristics of hydrologic features, the use of elevation models for identifying drainage paths can be influenced by flow barriers. To address these challenges, our focus in this study is on detecting drainage crossings through the application of advanced convolutional neural networks (CNNs). In pursuit of this goal, we use neural architecture search to automatically explore CNN models for identifying drainage crossings. Our approach not only attains high accuracy (over 97% for average precision) in object detection but also excels in efficiently inferring correct drainage crossings within a remarkably short time frame (0.268 ms). Furthermore, we perform a detailed profiling of our approach on GPU systems to analyze performance bottlenecks.  more » « less
Award ID(s):
1951741 2306184
NSF-PAR ID:
10473763
Author(s) / Creator(s):
; ; ; ; ;
Publisher / Repository:
ACM
Date Published:
Page Range / eLocation ID:
1780 to 1788
Subject(s) / Keyword(s):
["CNN inference throughput","GPU profiling","SPP-Net"]
Format(s):
Medium: X
Location:
Denver CO USA
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract

    Delineating accurate flowlines using digital elevation models is a critical step for overland flow modeling. However, extracting surface flowlines from high‐resolution digital elevation models (HRDEMs) can be biased, partly due to the absence of information on the locations of anthropogenic drainage structures (ADS) such as bridges and culverts. Without the ADS, the roads may act as “digital dams” that prevent accurate delineation of flowlines. However, it is unclear what variables for terrain‐based hydrologic modeling can be used to mitigate the effect of “digital dams.” This study assessed the impacts of ADS locations, spatial resolution, depression processing methods, and flow direction algorithms on hydrologic connectivity in an agrarian landscape of Nebraska. The assessment was conducted based on the offset distances between modeled drainage crossings and actual ADS on the road. Results suggested that: (a) stream burning in combination with the D8 or D‐Infinity flow direction algorithm is the best option for modeling surface flowlines from HRDEMs in an agrarian landscape; (b) increasing the HRDEM resolution was found significant for facilitating accurate drainage crossing near ADS locations; and (c) D8 and D‐Infinity flow direction algorithms resulted in similar patterns of drainage crossing at ADS locations. This research is expected to result in improved parameter settings for HRDEMs‐based hydrologic modeling.

     
    more » « less
  2. Abstract

    Drainage canals associated with logging and agriculture dry out organic soils in tropical peatlands, thereby threatening the viability of long‐term carbon stores due to increased emissions from decomposition, fire, and fluvial transport. In Southeast Asian peatlands, which have experienced decades of land use change, the exact extent and spatial distribution of drainage canals are unknown. This has prevented regional‐scale investigation of the relationships between drainage, land use, and carbon emissions. Here, we create the first regional map of drainage canals using high resolution satellite imagery and a convolutional neural network. We find that drainage is widespread—occurring in at least 65% of peatlands and across all land use types. Although previous estimates of peatland carbon emissions have relied on land use as a proxy for drainage, our maps show substantial variation in drainage density within land use types. Subsidence rates are 3.2 times larger in intensively drained areas than in non‐drained areas, highlighting the central role of drainage in mediating peat subsidence. Accounting for drainage canals was found to improve a subsidence prediction model by 30%, suggesting that canals contain information about subsidence not captured by land use alone. Thus, our data set can be used to improve subsidence and associated carbon emissions predictions in peatlands, and to target areas for hydrologic restoration.

     
    more » « less
  3. In the Baltimore urban long-term ecological research (LTER) project, (Baltimore Ecosystem Study, BES) we use the watershed approach to evaluate integrated ecosystem function. The LTER research is centered on the Gwynns Falls watershed, a 17,150 ha catchment that traverses a gradient from the urban core of Baltimore, through older urban residential (1900 - 1950) and suburban (1950- 1980) zones, rapidly suburbanizing areas and a rural/suburban fringe. Our long-term sampling network includes four longitudinal sampling sites along the Gwynns Falls as well as several small (40 - 100 ha) watersheds located within or near to the Gwynns Falls. The longitudinal sites provide data on water and nutrient fluxes in the different land use zones of the watershed (rural/suburban, rapidly suburbanizing, old suburban, urban core) and the small watersheds provide more focused data on specific land use areas (forest, agriculture, rural/suburban, urban). Each of the gaging sites is continuously monitored for discharge and is sampled weekly for chemistry. Additional chemical sampling is carried out in a supplemental set of sites to provide a greater range of land use. Weekly analyses includes nitrate, phosphate, total nitrogen, total phosphorus, chloride and sulfate, total suspended solids, turbidity, fecal coliforms, temperature, dissolved oxygen and pH. Cations, dissolved organic carbon and nitrogen and metals are measured on selected samples. This dataset presents stream chemistry from the Upper Gwynns Falls tributaries. From April 1999 to August 2000 Johns Hopkins University graduate student Mark Colosimo sampled a group of sites in the Upper Gwynns Falls (Red Run, Horsehead Branch, Scotts Level Branch, Holly Branch). There were two sites in the Red Run drainage. This watershed drains approximately 19 km2 and has been rapidly suburbanizing since the early 1990s. Percent impervious surface was approximately 10% as of 2002. Sampling station Red Run 1 (RR1) was approximately 35 m upstream of the crossing of Painters Mill Bridge Road, and 350 m upstream of the confluence with the Gwynns Falls. Sampling station Red Run 2 (RR2) was farther upstream, between the Pleasant Hill and Dolfield road crossings. There were two sites along Scotts Level Branch, an older suburban watershed which was approximately 25% impervious surface in 1970. Site SL1 drains approximately 11 km2 and is located at the outlet of the sub-watershed, just above the confluence with Gwynns Falls. Site SL2 is at the McDonogh Rd. bridge crossing. The Horsehead Branch (HH) sampling site was located at the McDonogh Road crossing. It drains approximately 5 km2 that has undergone rapid urbanization since the mid 1980s. As of 1997 percent impervious surface was approximately 12%. The Holly Bank (HB) sampling site was located just upstream of Gwynnbrook Ave. Seventy percent of land in this drainage is classified residential. The Gwynns Falls at McDonogh (GF5) site was located at the McDonogh school / McDonogh road crossing of the Gwynns Falls and samples a drainage area of approximately 51 km2, with approximately 20% impervious surface. 
    more » « less
  4. Assessment of lakes for their future potential to drain relied on the 2002/03 airborne Interferometric Synthetic Aperture Radar (IFSAR) Digital Surface Model (DSM) data for the western Arctic Coastal Plain in northern Alaska. Lakes were extracted from the IfSAR DSM using a slope derivative and manual correction (Jones et al., 2017). The vertical uncertainty for correctly detecting lake-based drainage gradients with the IfSAR DSM was defined by comparing surface elevation differences of several overlapping DSM tile edges. This comparison showed standard deviations of elevation between overlapping IfSAR tiles ranging from 0.0 to 0.6 meters (m). Thus, we chose a minimum height difference of 0.6 m to represent a detectable elevation gradient adjacent to a lake as being most likely to contribute to a rapid drainage event. This value is also in agreement with field verified estimates of the relative vertical accuracy (~0.5 m) of the DSM dataset around Utqiaġvik (formerly Barrow) (Manley et al., 2005) and the stated vertical RMSE (~1.0 m) of the DSM data (Intermap, 2010). Development of the potential lake drainage dataset involved several processing steps. First, lakes were classified as potential future drainage candidates if the difference between the elevation of the lake surface and the lowest elevation within a 100 m buffer of the lake shoreline exceeded our chosen threshold of 0.6 m. Next, we selected lakes with a minimum size of 10 ha to match the historic lake drainage dataset. We further filtered the dataset by selecting lakes estimated to have low hydrological connectivity based on relations between lake contributing area as determined for specific surficial geology types and presented in Jones et al. (2017). This was added to the future projection workflow to isolate the lake population that likely responds to changes in surface area driven largely by geomorphic change as opposed to differences in surface hydrology. Lakes within a basin with low to no hydrologic connectivity that had an elevation change gradient between the lake surface and surrounding landscape are considered likely locations to assess for future drainage potential. Further, the greater the elevation difference, the greater the drainage potential. This dataset provided a first-order estimate of lakes classified as being prone to future drainage. We further refined our assessment of potential drainage lakes by identifying the location of the point with the lowest elevation within the 100 m buffer of the lake shoreline and manually interpreted lakes to have a high drainage potential based on the location of the likely drainage point to known lake drainage pathways using circa 2002 orthophotography or more recent high resolution satellite imagery available for the Western Coastal Arctic Plain (WACP). Lakes classified as having a high drainage potential typically had the likely drainage location associated with one or more of the following: (1) an adjacent lake, (2) the cutbank of a river, (3) the ocean, (4) were located in an area with dense ice-wedge networks, (5) appeared to coincide with a potentially headward eroding stream, or (6) were associated with thermokarst lake shoreline processes in the moderate to high ground ice content terrain. We also added information on potential lake drainage pathways to the high potential drainage dataset by manually interpreting the landform associated with the likely drainage site to draw comparisons with the historic lake drainage dataset. 
    more » « less
  5. Abstract

    Unprecedented floods from extreme rainfall events worldwide emphasize the need for flood inundation mapping for floodplain management and risk reduction. Access to flood inundation maps and risk evaluation tools remains challenging in most parts of the world, particularly in rural regions, leading to decreased flood resilience. The use of hydraulic and hydrodynamic models in rural areas has been hindered by excessive data and computational requirements. In this study, we mapped the flood inundation in Huron Creek watershed, Michigan, USA for an extreme rainfall event (1000-year return period) that occurred in 2018 (Father’s Day Flood) using the Height Above Nearest Drainage (HAND) model and a synthetic rating curve developed from LIDAR DEM. We compared the flood inundation extent and depth modeled by the HAND with flood inundation characteristics predicted by two hydrodynamic models, viz., HEC-RAS 2D and SMS-SRH 2D. The flood discharge of the event was simulated using the HEC-HMS hydrologic model. Results suggest that, in different channel segments, the HAND model produces different degrees of concurrence in both flood inundation extent and depth when compared to the hydrodynamic models. The differences in flood inundation characteristics produced by the HAND model are primarily due to the uncertainties associated with optimal parameter estimation of the synthetic rating curve. Analyzing the differences between the HAND and hydrodynamic models also highlights the significance of terrain characteristics in model predictions. Based on the comparable predictive capability of the HAND model to map flood inundation areas during extreme rainfall events, we demonstrate the suitability of the HAND-based approach for mitigating flood risk in data-scarce, rural regions.

     
    more » « less