Using High‐Resolution Satellite Imagery and Deep Learning to Track Dynamic Seasonality in Small Water Bodies

Small water bodies (i.e., ponds; <0.01 km2) play an important role in Earth System processes, including carbon cycling and emissions of methane. Detection and monitoring of ponds using satellite imagery has been extremely difficult and many water maps are biased toward lakes (>0.01 km2). We leverage high‐resolution (3 m) optical satellite imagery from Planet Labs and deep learning methods to map seasonal changes in pond and lake areal extent across four regions in Alaska. Our water maps indicate that changes in open water extent over the snow‐free season are especially pronounced in ponds. To investigate potential impacts of seasonal changes in pond area on carbon emissions, we provide a case study of open water methane emission budgets using the new water maps. Our approach has widespread applications for water resources, habitat and land cover change assessments, wildlife management, risk assessments, and other biogeochemical modeling efforts.

imagery Wangchuk & Bolch, 2020). Many optical and active radar sensors used for water mapping do not have the requisite spatial or temporal resolutions to detect ponds and track the seasonality of their areal extent (Carroll & Loboda, 2017;DeVries et al., 2017;Du et al., 2019). Instead, past efforts have been biased toward lakes (Carroll & Loboda, 2017;DeVries et al., 2017;Pekel et al., 2016) and many water maps with high spatial resolutions are temporally static (Messager et al., 2016;Verpoorter et al., 2014). Undetected ponds introduce high uncertainty in estimations of CH 4 , as ponds tend to be hotspots for CH 4 production (Negandhi et al., 2013;Wik et al., 2016). Since existing methods often fail to detect these small water features, CH 4 budgets likely underestimate emissions of this potent greenhouse gas (Walter Anthony et al., 2018).
The prevalence and importance of small water bodies call for novel mapping approaches that utilize advancements in very high-resolution remote sensing. The capacities of commercial satellite platforms have expanded considerably in recent years, providing new opportunities for monitoring water bodies. Maxar's Quickbird, WorldView-2 and -3 platforms offer unparalleled image clarity at resolutions <0.5 m, allowing delineation of fine water features (e.g., Kaiser et al., 2021;Setiawan et al., 2019). However, Maxar imagery can be unaffordable for regional applications and change analysis. Planet Labs' PlanetScope constellation (Planet Team, 2021) has over 200 optical satellites with 3-5 m resolutions, daily revisits, and more affordable pricing. PlanetScope has previously been used to track inundation extent  and monitor the ecological variability of high latitude lakes (C. Kuhn et al., 2020. Research using PlanetScope to track lake shorelines found that open water mapping methods that rely on the normalized differenced water index (NDWI; McFeeters, 1996) ratios (i.e., higher pixel NDWI over water relative to dry ground) are unable to capture ponds in PlanetScope imagery ). An alternative water mapping method that used convolutional neural networks (CNN; i.e., deep learning) to detect lakes within the Himalaya using PlanetScope outperformed pixel-based approaches (Qayyum et al., 2020). However, the potential of PlanetScope imagery to track ponds remains unexplored.
This study leverages PlanetScope imagery to explore the feasibility of automated regional water body identification with CNNs, with the aim of detecting small ponds and seasonal changes in their extent. CNNs have proven effective in complex classification challenges, including detection of clouds (Caraballo-Vega et al., 2023;Kumthekar & Reddy, 2021), water flow paths (Stanislawski et al., 2021), and global real-time land-use/land-cover classification (Brown et al., 2022). We tested this approach using four case-study regions in Alaska and verified the water maps using on-ground Global Navigation Satellite System (GNSS) delineations. We then used monthly water body composite maps over the 3-year period to delineate areas having persistent open, standing water. Next, we investigated rapid surface water area variability by tracking individual water bodies throughout the 2021 snow-free season. Finally, we provided an example of how the new 3 m resolution water body maps compare to water maps from coarser-resolution satellite sensors, and a demonstration of how accounting for pond area greatly changes regional aquatic CH 4 budgets.

Study Areas
Our four study domains (i.e., areas of interest, AOI) were selected to span ecosystem gradients representative of boreal and tundra biomes in Alaska (Figure 1). Three AOIs are in Interior Alaska and represent complex boreal forest and wetland landscapes within a zone of discontinuous (50%-90%) permafrost (Pastick et al., 2015): Yukon Flats (YF; center coordinates: 66.509, −146.121; area: 9,440 km 2 ), Delta Junction (DJ; 63.826, −145.904; 7,110 km 2 ), and Fairbanks (FB; 64.793, −147.894; 6,200 km 2 ). DJ and FB cover terrains spanning headwaters within the Alaska Range to the low-lying Tanana River valley with numerous ponds and lakes. Both regions receive ∼210 mm of precipitation during the snow-free season (May-September). The YF is a low-lying fluvial plain with highly dynamic rivers, streams, lakes, and ponds Nitze et al., 2018;Rey et al., 2019), receiving 115 mm of summer precipitation. The fourth AOI is located within the Yukon Delta National Wildlife Refuge, which is part of the Yukon-Kuskokwim Delta (YKD; 61.177, −163.232; 7,690 km 2 ). This region is characterized by peat plateau and wetland tundra overlying sporadic (10%-50%) to discontinuous permafrost (Pastick et al., 2015). The YKD receives 310 mm of summer precipitation and water coverage ranges from isolated water bodies to high-limnicity zones having >15% surface water .

PlanetScope Imagery
We used Planet Labs PlanetScope orthotile surface reflectance imagery from the Dove-R and SuperDove sensors, provided through the NASA Commercial Smallsat Data Acquisition (CSDA) program. These two platforms provide daily coverage over the study regions from 2021 onward with less frequent coverage dating back to March 2019 (Roy et al., 2021). PlanetScope orthotiles are orthorectified, atmospherically corrected, and have a resolution of 3.125 m. The four-band imagery (blue, green, red, and near-infrared) overlaps closely with Sentinel-2 bands 2, 3, 4, and 8, allowing for testing and comparison of our deep learning method between PlanetScope and Sentinel-2. We used Planet's surface reflectance product to mitigate radiometric inconsistencies between scenes taken by different platforms at different viewing and solar zenith angles. Each PlanetScope orthotile has an accompanying "useable data mask" (UDM2) that flags pixels for snow, clouds, and haze (Planet Team, 2022).

Open Water Detection
This research focuses on detection of open, standing water, which we characterize as areas within lakes or ponds without obstruction from vegetation. We employ a deep learning approach (U-Net) to map open water as this approach considers spatial contextual information within the image and is likely less susceptible to uncertainties in the reflectance of individual pixels (Qayyum et al., 2020). Additional information on the U-Net approach and model development is provided in the Supporting Information S1 (Texts S1.1 and S1.2; Figures S1-S3).
The U-Net is a method of supervised classification, requiring training data. To train the U-Net (Text S1.3 in Supporting Information S1), we used PlanetScope orthotile imagery spanning a wide range of snow-free acquisition dates from the AOIs (Table S1 in Supporting Information S1 lists the training images). The training data set included 13,089 open water lakes and ponds that were hand-delineated in ArcGIS Pro v2.9 (ESRI, 2021), aided by 2 m Maxar WorldView-2/-3 imagery (Text S1.3 and Figure S2 in Supporting Information S1). We experimented with nine input data layers when training the U-Net. Input layers included PlanetScope red (R), green (G), blue (B), and near-infrared (NIR) surface reflectance bands. Additional candidate input layers included PlanetScope-derived: normalized difference vegetation index (NDVI); NDWI; saturation from the red, green, and blue channels (Sharma et al., 2015). NDVI, NDWI, and saturation layers were experimented with during initial model development, but were discarded from final model comparisons because they reduced accuracy scores considerably. To mitigate false-positive water detections in dark, terrain-shaded areas, we included layers derived from the 30 m National Elevation Dataset (NED) digital elevation model (3DEP DEM; U.S. Geological Survey, 2020) which describe slope angle and topographic location. Text S1.4 in Supporting Information S1 contains details on the derived model input layers. For model training, 70% of the delineated water body data set was used for training, 20% was used for validation during model training, and 10% was withheld for testing after training. The U-Net model was deployed over PlanetScope Dove-R and SuperDove imagery with <10% cloud cover over the 2019-2021 snow-free period (Table S2 in Supporting Information S1). Raw U-Net predictions were post-processed (Text S1.5 in Supporting Information S1) to develop final water maps for each image.

Classification Accuracy Assessments
Model accuracy was assessed using two individual test sets. The U-Net was first tested on the 10% withheld data. We assessed the water body predictions using common accuracy metrics for image classification and segmentation including precision, recall, F1, and Intersection over Union (IOU; See Text S1.6 in Supporting Information S1 for more details). Briefly, precision quantifies the model's ability to correctly label open water, recall quantifies the model's ability to detect all open water, F1 takes the harmonic mean between precision and recall, and IOU calculates the spatial overlap between training water masks and U-Net predictions. For further verification we used GNSS surveys of open water ponds to examine the U-Net's accuracy. The GNSS delineations were obtained by walking shorelines of six open water bodies in DJ, 46 in the YKD, and one in Fairbanks. IOU scores were calculated using the GNSS shorelines and corresponding U-Net predictions from PlanetScope imagery captured within 1 week of the surveys. See Text S1.7 in Supporting Information S1; Figures S4 and S5 in Supporting Information S1 for a detailed description of this process.

Composite Water Maps
Cloud cover resulted in loss of quality image retrievals. To generate high-confidence water maps with complete data coverage over the AOIs, we composited U-Net predictions from images that spanned the 3 years of Plan-etScope acquisitions (2019-2021). For each composite the occurrence ratio was calculated as the proportion of valid (cloud-free) observations where each pixel was detected as open water within a given time slice (Equation 1). Occurrence ratios ≥0.4 were given a value of 1 (open water), and occurrence ratios <0.4 were given a value of 0. This threshold was determined based on a visually optimal balance between false-positive (aggressive) and false-negative (conservative) predictions. 3-year composites were generated for each area using all available imagery from the snow-free periods. Monthly climatological composites were generated for DJ, FB, and YF AOIs for May-September using all imagery over the 3-year period for each given month. Monthly composites were not generated for the YKD due to lack of complete regional PlanetScope coverage for any given snow-free month. The monthly composites were averaged with the 3-year composites to further reduce geolocational error and anomalous predictions (Equation 2), which were more common in the raw monthly composites since they were generated with less overall predictions. (2)

Dynamic Water Tracking for 2021
The high cloud-free coverage provided by PlanetScope over each AOI in 2021 allowed us to effectively track individual water bodies over the May 24-September 19 period. First, individual water bodies were given unique identifiers by merging each monthly composite into a single map and adding a 30 m buffer around water bodies to avoid double-counting of single water bodies that were fragmented into multiple bodies by the model prediction . This buffered mask represents the maximum extent of all water bodies present in the composite water maps, allowing for area tracking of individual water bodies. Water bodies with less than five observations throughout the 2021 growing season and individual predictions that were greater than the maximum buffered extent were removed from the time series. We smoothed individual water body time series using a 7-observation median filter ( Figure S6 in Supporting Information S1) to mitigate differences in area retrievals due to variability in lighting and atmospheric conditions, and geometric distortions between different image acquisitions. For each water body we quantified changes in surface area with the total (Equation 3) and fractional dynamism (Equation 4; Cooley et al., 2019), where a is the area for a water body on any given day.

Product Comparisons
Water occurrence from the 3-year PlanetScope composites was compared with external water products derived from the Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat, and Sentinel-2 satellite constellations to assess the advantages of high spatiotemporal resolution PlanetScope imagery. . Water in rivers was manually masked out of these products, and an occurrence ratio of 40% was used to threshold the MOD44W and JRC SWO maps. Text S2 in Supporting Information S1 contains details on the derivation of these products. A further step was to evaluate the Planet-derived maps against those derived from Sentinel-2. Given the spectral overlap of the PlanetScope blue, green, red, and near-infrared with Sentinel-2 bands 2, 3, 4, and 8, it was hypothesized that our U-Net approach might also be useful for detecting water within Sentinel-2 imagery. To test this hypothesis, we developed 3-year (2019-2021) Sentinel-2 composite water maps using the Planet-trained U-Net (See Text S2.3; Figure S7 in Supporting Information S1).

Water Map Application: Upscaling Methane Emissions
We used our resulting Planet-derived 3 m water maps to investigate the potential impact on high latitude CH 4 emission budgets, when accounting for ponds (also see Text S3 in Supporting Information S1). For this we used CH 4 fluxes from the Boreal-Arctic Wetland and Lake Methane Data set (BAWLD-CH 4 ; Kuhn, Varner, et al., 2021) to upscale emission estimates across each AOI. BAWLD-CH 4 reports median CH 4 emissions for different size classes of water bodies (small, medium, large) based on observations of CH 4 diffusion (1,134 lakes and ponds) and ebullition (202 lakes and ponds). We multiplied these median CH 4 emissions by the total area of each water body size class. In addition, we explored the co-location of spatial patterns observed between the 3 m water maps and >21,000 CH 4 hotspots in the YKD determined from hyperspectral Airborne Visible-Infrared Imaging Spectrometer -Next Generation (AVIRIS-NG) in an approach similar to Baskaran et al. (2022) and Elder et al. (2020Elder et al. ( , 2021.

Water Map Accuracy Assessments and Detection Limits
The U-Net approach detected ponds and lakes in our AOIs with high accuracy when using PlanetScope RGB NIR bands and 3DEP DEM-derived slope angle as inputs. Addition of the topographic layers (slope angle and TWI) had a negligible impact on accuracy scores, however, the slope layer greatly reduced false-positive detections over steeply sloping terrain based on visual assessment of the water maps. The trained U-Net had respective IOU, precision, recall and F1 scores of 0.91, 0.96, 0.95, and 0.95 (Text S1.8; Table S3 in Supporting Information S1). These scores are comparable to previous U-Net water body classifiers that used PlanetScope to map glacial lakes (Qayyum et al., 2020). Using the GNSS-based delineations (Text S1.7; Figure S5 in Supporting Information S1) the average IOU for ponds within the >0.001 km 2 and <0.01 km 2 size range was 0.81. Ponds within a smaller (>0.0001 km 2 and <0.001 km 2 ) size range had an average IOU of 0.40; 54% of these were at least partially detected. Based on this ground-surveyed data set we found the smallest detectable water body area 10.1029/2022GL102327 6 of 12 to be 0.0001 km 2 (Text S1.7; Figure S5 in Supporting Information S1), which is approximately equal to nine PlanetScope pixels. This size threshold is slightly higher than the minimum detectable water body size reported by studies using 2 m TerraSAR X radar (Muster et al., 2013) and <2 m Maxar optical imagery (Bouchard et al., 2014). Our method's detection limit is one order of magnitude finer than the Global Water Bodies Database (GLOWABO; Verpoorter et al., 2014) based on 15 m pan-sharpened Landsat imagery, and three orders of magnitude finer than the HydroLakes global lake data set (Messager et al., 2016).

Regional Surface Water Characteristics
The 3 m PlanetScope water maps allowed us to track ponds, in addition to lake areas, across the four AOIs ( Figures S8a-S8l in Supporting Information S1, Movies S1-S4). Of the 52,797 water bodies present in the 3-year PlanetScope composite water maps, 77% were ponds, which had considerably higher percent dynamism than lakes over the 2021 snow-free period ( Figure S9 in Supporting Information S1). Despite considerable seasonal changes in lake and pond surface area (Figures 2a-2d), monthly water body areas follow similar size distributions (Figures 2e-2h) found by previous studies (Kyzivat et al., 2019;Messager et al., 2016;Seekell & Pace, 2011;Verpoorter et al., 2014). Although size distributions did not change considerably across AOIs and months, the PlanetScope water body maps and area tracking highlight distinct differences in spatiotemporal water body characteristics between the four AOIs. Lakes and ponds in DJ and FB had similar spatiotemporal characteristics due to the proximity of these two AOIs. The 3,481 detected ponds in DJ made up 21%-24% of the total lake and pond surface area in this AOI based on the monthly PlanetScope composites (Figures S10a and S10b in Supporting Information S1). Total pond surface area decreased by 31% in DJ during the 2021 snow-free season (Figure 2) which contributed to nearly half (43%) of total dynamism in water body surface area (Figures S10c and S10d in Supporting Information S1). DJ contained many stable lakes focused in lowland areas surrounding the Tanana River and its tributaries. Ponds had a greater contribution to total water body area and dynamism in FB due to their relative abundance, with a notable area south of the Tanana River where topography is extremely flat and the landscape is scattered with many small ponds that showed rapid decreases in surface area during summer. The monthly composite maps indicate that a total of 4,011 ponds in FB accounted for 26%-28% of total lake and pond surface area (Figures S10a and S10b in Supporting Information S1). The total surface area of ponds in FB decreased by 29% in 2021 (Figure 2), which represented 60% of the total dynamism in lake and pond surface area (Figures S10c and S10d in Supporting Information S1). Seasonal declines in lake and pond surface area in these AOIs contrasts with decadal-scale assessments that show increases in pond size and number due to melting permafrost (Walter Anthony et al., 2021). The dense time series imagery provided by PlanetScope can therefore be leveraged to decouple seasonal from climatological changes in water body characteristics.
In the YF, previous studies have found many large lakes with high dynamism . It is hypothesized that this dynamism is potentially due to the connectivity of these lakes to groundwater beneath the permafrost (Rey et al., 2019), which is associated with increased hydraulic conductivity (Walvoord et al., 2012). The PlanetScope maps showed water bodies in YF to be larger on average (0.04 km 2 ) than DJ (0.01 km 2 ) and FB (0.006 km 2 ) (Figures S8g-S8i in Supporting Information S1). Given the abundance of lakes in YF it is unsurprising that the 6,626 ponds made up only 5% of total water body area in this AOI (Figures S10a and S10b in Supporting Information S1). Ponds experienced a 40% decrease in total surface area during 2021 (Figure 2) which represented 22% of the total dynamism in water body surface area (Figures S10c and S10d in Supporting Information S1). In YF, ponds contributed less to total surface water variability than DJ and FB due to the large number of lakes with high percent dynamism ( Figure S9 in Supporting Information S1).
Water body coverage was much more spatially heterogeneous in the tundra landscape of the YKD, ranging from zones of isolated ponds in the center of the AOI, zones of high limnicity with large, densely packed lakes in the south, and a small area in the northwest of the AOI with little to no water bodies ( Figure S8j in Supporting Information S1). The YKD had a much higher percentage of open water bodies (27.9%) than the boreal AOIs (DJ: 0.6%, FB: 0.4%, YF: 4%) and water bodies were larger on average (0.06 km 2 ). A total of 26,624 ponds detected in the YKD made up 5% of the total water body area in the YKD (Figures S10a and S10b in Supporting Information S1). Similar to the boreal AOIs, the variability of pond surface area clearly outsized the contribution of ponds to total water body area. However, surface area increased for both lakes (2%) and ponds (18%) in the YKD in 2021 (Figure 2), and ponds contributed to 10% of total water body surface area variability (Figures S10c and S10d in Supporting Information S1). During July 2021, the YKD experienced the greatest amount of precipitation for July on record since the 1920s (NOAA National Centers for Environmental Information, 2021), indicating that the observed variability in water body surface area in the YKD could be rainfall driven. This variability contrasts with the boreal regions which showed characteristics of snowmelt-driven change. In addition to rainfall, permafrost thaw in the YKD could also have impacted surface water area over the 2021 summer (Grosse et al., 2013). The seasonality of water body surface area in the YKD is poorly characterized in previous research, however, studies have found late-season increases in lake area in the Mackenzie River Delta  and water area increases across multi-year timescales in Arctic deltas (Carroll & Loboda, 2017;Nitze et al., 2018).

Evaluation Against Other Surface Water Products
The PlanetScope water body maps provided insights into the importance of capturing ponds for evaluation of seasonal water body characteristics. We compared the PlanetScope water body maps with water maps from coarser-resolution remote sensing imagery to investigate the advantages of the high-resolution commercial imagery ( Figures S8a-S8l in Supporting Information S1). Statistics from the comparisons are summarized in Table S4 and Figures S11a-S11d in Supporting Information S1.
Over the four AOIs, the 3 m PlanetScope water maps detected 8% more total water body surface area than the 10 m Sentinel-2 product, 18% more surface area than the 30 m JRC SWO (Pekel et al., 2016), and 22% more surface area than the 250 m MOD44W . Discrepancies between the PlanetScope water maps and the coarser-resolution products were almost entirely based on water body surface area relative to the spatial resolution of each product. IOU scores between each respective product and the PlanetScope maps diminished with decreasing water body surface area ( Figure S11a in Supporting Information S1), approaching zero at different area thresholds for each product (Figures S11c and S11d in Supporting Information S1). MOD44W failed to detect 44,000 water bodies, as the 250 m MODIS pixels largely limit detection to water bodies >0.1 km 2 . 38,000 water bodies present in the PlanetScope water maps were missing from the JRC SWO product, which is likely due to the difficulty in resolving water bodies <0.01 km 2 with the 30 m Landsat resolution. The 10 m Sentinel-2 product captured water bodies as small as 0.001 km 2 , but this lower limit still resulted in 23,000 missed water bodies that were present in the PlanetScope water maps.
Temporal differences between the water body products led to disagreement in the YKD, where the coarser-resolution products failed to detect many larger water bodies ( Figure S11d in Supporting Information S1). Inspection of the PlanetScope 3-year composite water body maps and PlanetScope imagery indicates that many of these water bodies were only present for brief periods of time after snowmelt (Late May). Since cloud-free observation windows are sparse over the YKD, the absence of these rapidly diminishing water bodies in the three external products was likely due to their relatively coarse temporal resolutions with respect to PlanetScope.
In terms of total water body area, differences in water detection capabilities between PlanetScope and the highest-resolution product that was compared (Sentinel-2; 10 m) seem minimal when considering the costs of commercial satellite data. However, PlanetScope's advantage clearly lies in its ability to detect small ponds >0.0001 km 2 , which resulted in 77% more water bodies represented in the PlanetScope maps than in the Sentinel-2 maps. Additionally, the daily revisit of PlanetScope provides five times the amount of observations as the Sentinel-2a -2b constellation (5-day revisit), allowing for more observations within cloud-free windows that are sparse in many regions of the globe.

Tracking CH 4 Emissions From Open Water and CH 4 Hotspot Occurrence
CH 4 from ponds has largely been excluded from regional CH 4 budget studies due to the difficulty in detecting ponds using coarser-resolution water body datasets. Through spatial upscaling of aquatic flux rates based on the BAWLD-CH 4 synthesis (C. Kuhn, John, et al., 2021) and our new 3 m PlanetScope open water maps, we estimated that ponds and lakes contributed to 0.39 (DJ), 0.35 (FB), 1.57 (YF), and 6.26 (YKD) mg CH 4 per day per m 2 of each respective AOI. Considering the temporal variability in pond extent based on the monthly PlanetScope composites, ponds alone contributed to 24%-28% of upscaled aquatic CH 4 emissions in DJ, 30%-37% in FB, 8%-10% in YF, and 15% in the study portion of the YKD ( Figure S12 in Supporting Information S1). The observed temporal variability in CH 4 emissions from lakes and ponds is conservative compared to PlanetScope-based estimates from the mid-latitudes (Hondula et al., 2021), which could be due to environmental factors or the relatively coarse temporal resolution of our monthly composite water maps. Additionally, comparison of the AVIRIS-NG-derived CH 4 hotspot maps to the 3-year PlanetScope composite revealed that CH 4 hotspot occurrences were three times higher around ponds (177 hotspots per km 2 ) than lakes in the study portion of the YKD (54 hotspots per km 2 ) (Text S3.2, Figure S13 in Supporting Information S1). Given the detection limits of the coarser-resolution surface water products investigated in this study, the contribution of ponds to total upscaled aquatic CH 4 emissions would not be quantifiable with Landsat and MODIS-based surface water products. Although Sentinel-2 showed the ability to detect ponds, total upscaled CH 4 emissions from ponds were still 28% lower over the snow-free season when using the Sentinel-2 product as opposed to PlanetScope ( Figure S12 in Supporting Information S1), due to the ability of PlanetScope to capture ponds an order of magnitude smaller than Sentinel-2.
Area alone is not always a good predictor of CH 4 emissions from water bodies (Kohnert et al., 2018;Rehder et al., 2021), the emission data set used in our upscaling is not necessarily representative of aquatic CH 4 emissions in our AOIs, and the PlanetScope water maps generated in this study do not identify wet vegetated areas that tend to produce higher CH 4 emissions (Kyzivat et al., 2022;Villa et al., 2021). However this example presents the first critical steps in incorporating the effects of ponds and their seasonality on regional aquatic CH 4 emissions and CH 4 hotspot occurrence.

Conclusions
High spatiotemporal resolution satellite imagery from Planet Labs and deep learning provide new and improved means for detecting and tracking very small water bodies over time. These water maps have applications across many disciplines that require monitoring rapid changes in surface water extent, and modeling applications where small water bodies have a considerable influence on the relevant process (e.g., CH 4 emissions). There are undoubtedly ponds <0.0001 km 2 that our PlanetScope imagery is unable to resolve. While these small water bodies may not contribute significantly to total surface water area and variability, they likely have a much greater influence on the total number of water bodies and CH 4 emissions. With ever-improving capabilities of commercial optical and new radar satellite platforms (e.g., Capella Space X-band synthetic aperture radar), the capacity for high spatiotemporal resolution monitoring of surface water is greater than ever. We present an approach that leverages these new capabilities to enhance our understanding of fine-scale land surface changes that have broad-scale implications.