https://orcid.org/0000-0001-7402-2825
Figures
Abstract
Site-specific agricultural water management is particularly important in highly productive and diverse agricultural regions such as California’s Santa Clara Valley (SCV), where broad crop categories can obscure substantial parcel-scale differences in consumptive use. This study used unsupervised machine learning on remotely sensed data from 2019-2023 to develop operationally distinct, agricultural consumptive-use groups for major crop types. Time series of Sentinel-2 normalized difference vegetation index (NDVI), OpenET ensemble actual evapotranspiration (ETa, used as a proxy for consumptive use), and PRISM precipitation were analyzed at the parcel level for truck crops, vineyards, and hay crops. Across 2,189 parcels (~7,483 ha), the only-NDVI and NDVI+ETa clustering approaches identified 13 consumptive-use clusters and revealed substantial within-crop heterogeneity hidden by conventional crop averages. Six clusters were detected in truck crops (TC), four in vineyards (VC), and three in hay crops (HC). Adding ETa magnitude and trend information generally reduced average within-cluster coefficient of variation (CV) relative to only-NDVI clustering and increased separation among cluster means. The ratios of between-cluster to mean within-cluster CV were lower under only-NDVI (TC = 0.30, VC = 0.46, HC = 0.74) than under NDVI+ETa (TC = 1.35, VC = 1.06, HC = 1.07), compared with crop-wide CVs of 17%, 24%, and 20%, respectively. Truck crops showed the strongest seasonal and interannual heterogeneity, vineyards exhibited clearer spatial differentiation and a modest decline in consumptive use, and hay crops separated into distinct pasture and grain-forage systems with contrasting delivery needs and climate sensitivity. Across all three crop types, observed consumptive-use patterns were consistent with hydroclimatic stresses and management practices, including cover cropping, irrigation technology, deficit irrigation, and seasonal fallowing, although direct attribution would require independent field, permit, or metering data. These findings support customized agricultural water-management plans for irrigation monitoring, conservation planning, and adaptive water management in groundwater-dependent regions.
Citation: Sarwar A, Medellín-Azuara J, Abatzoglou JT, Viers JH (2026) Identifying agricultural consumptive-use patterns to support adaptive water management in California’s Santa Clara Valley via remote sensing and machine learning. PLOS Water 5(7): e0000416. https://doi.org/10.1371/journal.pwat.0000416
Editor: Tafadzwanashe Mabhaudhi, London School of Hygiene and Tropical Medicine / University of KwaZulu-Natal School of Agricultural Earth and Environmental Sciences, UNITED KINGDOM OF GREAT BRITAIN AND NORTHERN IRELAND
Received: January 21, 2025; Accepted: July 10, 2026; Published: July 29, 2026
Copyright: © 2026 Sarwar et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The Crop water use datasets are available at the OpenET (https://etdata.org/) a collaborative led by NASA, the Desert Research Institute (DRI), and the Environmental Defense Fund (EDF), with support from Google Earth Engine at https://developers.google.com/earth-engine/datasets/catalog/OpenET_ENSEMBLE_CONUS_GRIDMET_MONTHLY_v2_0. Satellite datasets for the surface reflectance from the Sentinel 2 by the European Space Agency (ESA) were processed at the Google Earth Engine available at https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2_SR_HARMONIZED. Crop types were gathered from the LANDIQ (https://www.landiq.com/) developed by Land IQ, LLC for the California Department of Water Resources available at https://data.cnra.ca.gov/dataset/statewide-crop-mapping. Precipitation datasets were downloaded from the Google Earth Engine available at https://developers.google.com/earth-engine/datasets/catalog/OREGONSTATE_PRISM_AN81m provided by the PRISM Climate Group, Oregon State University. County crop-type datasets are catalogued at the County of Santa Clara Division of Agriculture website at https://ag.sccgov.org/crop-reports-newsletters-monthly-agricultural-updates. All the datasets used are available the following link https://ucmerced.box.com/s/q87dssfehvggk1ehotbk07he46ymki52.
Funding: This work was supported in part by the Santa Clara Valley Water Agency (Valley Water), National Science Foundation (NSF)-United States Department of Agriculture (USDA) National Institute of Food and Agriculture (NIFA) under the AI Research Institutes program: Agricultural AI for Transforming Workforce and Decision Support (AgAID) (Award No. 2021-67021-35344 to AS, JMA, JTA, and JHV), and by the USDA-NIFA Agriculture and Food Research Initiative: Securing a Climate Resilient Water Future (SWF) (Award No. 2021-69012-35916 to JMA, JTA, and JHV). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
Irrigated agriculture accounts for approximately 70% of global freshwater withdrawals and 40% of food production from 20% of farmland [1]. Climate change is expected to increase irrigation water requirements by 3–5% in the 2020s and by 5–8% by the 2070s relative to 1961–1990 [2]. Reduced water availability can threaten sustainable food production in irrigation-dependent agricultural systems, including South Asia, the Mediterranean, and the western United States [3]. Groundwater-dependent agricultural regions such as California’s Santa Clara Valley and Central Valley therefore require more precise accounting of crop consumptive use as water demand rises and groundwater resources decline [4].
Water-balance accounting is central to agricultural water management and irrigation scheduling [5–7]. Field instruments such as soil-moisture sensors [8,9], lysimeters [10], and atmometers [8] can monitor evapotranspiration (ET) [11] and soil-water deficit [12], while high-efficiency irrigation methods [13,14] and scheduling tools [10] can reduce applied water and improve consumptive-use efficiency. These approaches are valuable but often require dense instrumentation, expert supervision, and farm-level data. Scalable, non-invasive methods are therefore needed to complement field monitoring and support district-level irrigation accounting across many parcels.
Recent advances in agricultural system science and engineering combine domain-specific water problems with new intelligence tools [15]. Machine learning (ML) and remote sensing (RS) have expanded geospatial analysis of earth observations, including agricultural water budgets [7]. Cloud-based platforms such as Google Earth Engine (GEE) provide scalable tools for processing large spatiotemporal datasets [16]. Within this setting, unsupervised clustering can help identify groups of fields with similar crop phenology and consumptive-use behavior. Clustering has long been used in precision agriculture [17], land-cover mapping [18,19], and vineyard management [20,21], and may therefore help distinguish agricultural water-use patterns that are missed by crop averages.
This study examines whether unsupervised clustering can differentiate agricultural consumptive-use groups in the Santa Clara Valley (SCV) of California, where actual evapotranspiration (ETa) is treated as a proxy for consumptive use. The SCV is groundwater-dependent for irrigation despite being located within a developed water-delivery district, and it supports 22 specialty crops [22]. The Santa Clara Valley Water District (hereafter, Valley Water) monitors, measures, and prices water in this region [23]. Valley Water currently uses metered agricultural-well data together with a self-reported Table of Averages to estimate agricultural consumptive use. This worksheet-based approach combines crop type, cultivated area, and estimated consumptive use. Although roughly 65% of irrigated acreage is metered, the long tail of small, unmetered users and mixed operations limits parcel-level inference. District managers also recognize that irrigation technologies and conservation measures can create high variability in applied water and actual consumptive use within a crop type [24]. Even where estimated irrigation efficiencies exceed 90% [25], mixed truck-crop and nursery production remains difficult to quantify with standardized crop averages.
This paper presents unsupervised time-series clustering of Normalized Difference Vegetation Index (NDVI) data, combined with satellite-derived actual evapotranspiration (ETa), to identify discrete agricultural consumptive-use groups. To our knowledge, this is the first parcel-scale application in a groundwater-dependent specialty-crop region that combines Sentinel-2 phenological clustering with OpenET-derived ETa to define operational consumptive-use strata for district water accounting and adaptive allocation. The analysis provides an alternative to methods that rely only on crop-specific coefficients or average consumption tables. Specifically, the study asks: (1) How large are consumptive-use variations within the major agricultural crop categories of the SCV relative to the Table of Averages approach? (2) Can distinct consumptive-use groups be identified within a single crop category using cloud-based geospatial data and unsupervised clustering?
2. Materials and methods
An unsupervised time-series clustering workflow was applied to parcel-scale agricultural irrigation consumptive use in the SCV. The workflow used readily available remote sensing datasets and on-demand cloud computing to distinguish operational consumptive-use groups in a repeatable and scalable manner.
2.1. Santa clara valley use case
The Santa Clara Valley (SCV) is situated south of San Francisco Bay, featuring a matrix of highly urbanized areas and intensive agriculture (Fig 1A). The SCV region is characterized by a Mediterranean climate with warm, dry summers and mild, wet winters, which provides an ideal environment for a wide range of high-value crops, including fresh produce (e.g., leafy greens, mushrooms, peppers, cole crops), winegrape vineyards, and nursery crops [22,26]. The SCV’s average annual precipitation reflects its topography and coastal proximity, ranging from 1,346 mm in the Santa Cruz Mountains to 355 mm on the valley floor. This geographic variability in precipitation is set against an average annual crop water requirement of 800 mm [27] as shown in Fig 1B and Fig 1C.
Inset (B) shows the mean monthly precipitation in mm along with the mean monthly minimum and maximum temperatures (°C) for the valley from 1991 to 2020. (C) shows the SCV crop type distribution by percent area.
More advanced agricultural practices, such as deficit irrigation, cover crops, early-season irrigation, and multiple cropping, are employed on half of the region’s agricultural parcels, which are predominantly dominated by annual crops and typically small in size (i.e., 4 hectares or less) [22]. Major irrigation technologies employed in the SCV include sprinklers and micro-sprinklers, primarily for perennials, and drip irrigation (either surface or subsurface) for annual crops. Single cropping and double cropping are practiced in the SCV, with vegetables and leafy greens commonly grown together, or as a cover crop between the rows in a vineyard [24]. In contrast to other Californian agricultural areas, where permanent fruit and nut crops are gradually replacing truck and field crops [28], also, most commodities produced in Santa Clara County have remained relatively consistent over the last ten years [25].
2.2. Data sources
The clustering approach used publicly available online datasets to support a workflow that can be repeated by water agencies and researchers. The data sources are described below by type.
2.2.1. Land use data.
Crop types at the parcel scale were obtained from Land IQ, LLC (LandIQ) land-use surveys prepared for the California Department of Water Resources [29]. LandIQ is generally more reliable than alternative products for California specialty crops [30]. For context, the LandIQ statewide product sequence includes maps for 2014 and 2016, followed by annual water-year mapping beginning in 2018; a 2015 statewide LandIQ map is not part of this product sequence. This study did not combine multiple LandIQ crop-map years. Instead, the 2020 water-year LandIQ map was used as the sole crop-type layer and fixed spatial reference for extracting ETa and NDVI across water years 2019–2023. Using one reference year kept parcel boundaries and crop labels consistent across the multi-year remote-sensing window, while the resulting limitation from rotations or crop changes is addressed in the Limitations section. The product categorizes nearly 6.2 million ha of land use into more than 50 crop and land-use types, with reported accuracies exceeding 98% across more than 440,000 individually classified polygons and minimum field sizes of 0.20-0.80 ha (0.5-2.0 acres), depending on crop type [29,31].
2.2.2. Evapotranspiration data.
OpenET provides monthly actual evapotranspiration (ETa) data for the western United States [32]. This study used the OpenET ensemble, which combines six satellite-driven energy-balance and water-balance models at 30 m spatial resolution. ETa data were extracted for each LandIQ crop type for water years 2019–2023 using the centroid of each SCV agricultural parcel [32].
2.2.3. Precipitation data.
High-resolution geospatial climate records for the United States were obtained from PRISM (Parameter-elevation Regressions on Independent Slopes Model) [33,34]. For the WYs 2019–2023, PRISM data, including temperature, humidity, and precipitation (PPT), were extracted for field centroids.
2.2.4. Satellite datasets.
European Space Agency Sentinel-2, atmospherically corrected surface reflectance, Level-2A, at 10 m spatial resolution and a 5-day revisit interval, were processed using GEE services [35].
2.3. Data preprocessing and clustering technique
Remotely sensed data from satellite imaging platforms have been widely used in agricultural studies, and Sentinel-2 products are increasingly used to assess crop health [36] and irrigation water use [37] because of their high spatial and temporal resolution. These high-resolution datasets are prone to cloud cover and atmospheric effects that can affect accuracy. To reduce these random errors [38], a Savitzky-Golay (SG) moving-window filter was used to fill spatiotemporal gaps and smooth Sentinel-2 NDVI time series after resampling to a 5-day, 30 m product [39].
Unsupervised clustering was used to distinguish crop-specific agricultural consumptive-use groups. Several clustering methods are available for time-series problems, including k-means [40,41], fuzzy c-means [42], k-medoids [43], and fuzzy c-medoids [44]. Few studies [45–47] have applied these approaches to spatiotemporal partitioning of consumptive-use clusters, and none has done so for the SCV region.
Two complementary clustering views were retained. The only-NDVI approach provides a scalable phenology-first grouping based on optical imagery alone. This matters for transferability because reliable ETa products may be unavailable, delayed, regionally inconsistent, or model-dependent in other agricultural regions, whereas NDVI can often be generated from widely available satellite imagery. The only-NDVI workflow therefore provides a practical first-pass classification and an ET-independent benchmark. The combined NDVI+ETa approach was used where reliable ETa estimates were available to add consumptive-use magnitude and interannual trend features and to test whether phenological groups also corresponded to meaningful consumptive-use differences (Text B, Text C, and Table D to Table I in S2 Appendix). Parcel-based mean NDVI time series were used to develop clusters within each selected major crop category. Absolute NDVI describes the canopy condition of a parcel through time, whereas correlation-based similarity groups parcels whose NDVI profiles rise and fall together even when their absolute values differ. A dissimilarity matrix was therefore calculated from pairwise time-series differences, and the Partitioning Around Medoids (PAM) k-medoids algorithm was applied to this matrix [48]. PAM minimizes dissimilarity between observations and representative medoids, tolerates outliers better than k-means, and reduces the influence of noisy observations, although it is computationally slower than k-means [49].
Candidate cluster numbers from k = 2 to k = 8 were evaluated before selecting the final number of clusters. For the only-NDVI workflow, the grouping structure was selected from the internal global pseudo F-statistic generated by the time-series clustering tool and checked against interpretability. For the NDVI+ETa workflow, ETa-based prediction error, effect size, cohesion/separation metrics, and interpretability were used to balance practical parsimony against marginal predictive gain. Larger k values often continued to improve prediction error at diminishing returns; the selected k values preserve operational utility for water managers while retaining most explanatory gain. ETa and ETa-derived features were used only in the NDVI+ETa clustering approach; Text A, and Table A to Table C in S2 Appendix, reports the cluster-selection statistics. Clusters are labeled by crop category in descending order of area: truck crop clusters TC1-TC6, vineyard clusters VC1-VC4, and hay crop clusters HC1-HC3. Dataset preprocessing, modeling, and analysis were carried out in GEE, Python v3.11, and ArcGIS Pro (ESRI v3.4).
2.4. Evaluation metrics
Following Calinski and Harabasz [50], the pseudo F-statistic was calculated as the ratio of between-cluster variance to within-cluster variance. Candidate k values from 2 to 8 were compared using pseudo F-statistic, separation metrics (silhouette and Davies-Bouldin), and parcel-level cross-validated ETa error.
where k is the number of clusters and n is the number of observations at any clustering step.
3. Results
Overall, the results describe agricultural consumptive-use groups that emerged within each major crop category from two complementary clustering approaches: K-medoid clustering of NDVI correlation profiles (only-NDVI) and combined clustering based on NDVI+ETa features. The central pattern is not simply that truck crops, vineyards, and hay crops use different total water volumes. Rather, each broad crop category contains multiple parcel-scale consumptive-use groups whose seasonal timing, magnitude, and interannual responsiveness reflect both hydroclimatic conditions and farm management practices.
3.1. Clustering analysis
Clusters (Sections 2.3-2.4) were created separately for truck crops, vineyards, and hay crops because these agricultural systems present different irrigation and water-accounting concerns. Candidate k values were evaluated with cohesion/separation metrics, parcel-level ETa prediction error, effect size, sampling-reduction potential, and interpretability. The only-NDVI approach was retained as a scalable, ET-independent phenology classification that can be applied where reliable ETa data are not yet available or are not sufficiently trusted for management use. In this study, the NDVI+ETa approach served as the complementary partition that incorporates consumptive-use magnitude and trend once reliable ETa estimates are available. Detailed number-of-cluster selection statistics for NDVI+ETa are provided in Text A, and Table A to Table C in S2 Appendix.
For truck crops, six clusters were selected. Parcel k-fold RMSE and effect size (omega-squared) improved through k = 6 before leveling off, cohesion/separation metrics remained acceptable, sampling-size reduction relative to simple random sampling exceeded 50%, and MAPE reached about 10%. Four clusters were selected for vineyards. RMSE and omega-squared improved substantially from k = 1 to k = 4, while additional clusters produced only modest gains; k = 4 reduced k-fold RMSE by about 27.2%, whereas k = 7–8 improved it only to about 29.9-30.2%. Three clusters were selected for hay crops because k = 3 was the lowest-complexity solution at the main elbow in parcel k-fold RMSE, effect size, cohesion/separation metrics, and sampling reduction. Larger k values continued to improve some statistics, but the added complexity was not justified for operational interpretation.
The all-parcel versus cluster-mean CV comparison is descriptive rather than a formal statistical test because these metrics summarize different quantities. Cluster usefulness was therefore evaluated using ETa prediction error, MAPE, omega-squared, within-cluster CV, and the complementary NDVI+ETa partition reported in S2 Appendix. Cross-crop checks show that broad crop means hide substantial parcel-scale variation: all-parcel CVs are about 17% for truck crops, 24% for vineyards, and 20% for hay crops, compared with only-NDVI CVs across cluster means of about 5%, 10%, and 12%, respectively (S1 Appendix). The NDVI+ETa comparison further shows that adding ETa widens between-cluster separation and generally lowers average within-cluster CV relative to only-NDVI (Table F in S2 Appendix). These results indicate that k = 2 or crop means alone would obscure important agricultural water-use patterns (Fig 2).
3.1.1. Truck crops.
A total of six truck-crop clusters (TC) emerged from more than 30 truck-crop subclasses and 1,044 parcels in the SCV study area. Significant differentiation in consumptive use and seasonal NDVI signals was detected among the six clusters (Fig 3, Fig 5, and S2 Appendix). The truck-crop cluster mean annual consumptive-use CV was 6%, ranging from 5% in 2019 to 8% in 2023 (Fig 3). The same 1,044 truck-crop parcels yield six area-ranked clusters under either method, although labels are not directly interchangeable unless crosswalked. Under only-NDVI, cluster means of consumptive agricultural water use span a narrow 3,116-3,492 mm range (1.1-fold), and within-cluster CVs of 10–22% sit close to the crop-type CV of 17%. This grouping captures seasonal timing more strongly than absolute consumptive-use magnitude. NDVI+ETa stretches the cluster-mean range to 2,654-4,224 mm (1.6-fold) and reduces every within-cluster CV below the crop-type level (8–17%). The methods agree on TC1: both versions identify a mainstream irrigated-vegetable group near 3,500 mm. They differ most at TC6. NDVI+ETa identifies a 4,224 mm year-round-canopy subgroup with a tight CV of 12%, whereas only-NDVI TC6 remains near the crop-type mean and does not resolve this magnitude extreme. Within the NDVI+ETa scheme, TC1 is the tightest cluster and peaks in August; TC6 remains green year-round; TC2 and TC4 have similar mean ETa but peak in April and June, respectively; TC3 has the steepest declining ETa trend (-24 mm/yr); and TC5 has near-zero canopy despite substantial consumptive use, consistent with greenhouse or bare-soil parcels. These contrasts show that only-NDVI is useful as a scalable phenological first pass, while NDVI+ETa exposes high- and low-volume subgroups.
The inset plot is the coefficient of variation.
As reported in Table 1 and Text B, Text C, and Table D to Table I in S2 Appendix, the within-cluster and between-cluster statistics indicate significant differences in total consumptive-use distributions.
As shown in the Sankey diagram, Fig 2, TC1 has by far the largest planted area (i.e., 1536 ha) and TC6 the smallest (~151 ha). TC1 and TC2 are the most diverse and are mainly grown to miscellaneous truck vegetables, whereas TC3–TC5 lean toward flowers/nursery crops, and TC6 has smaller mixed areas including melons and berries. Across all clusters, miscellaneous truck crops occupy the most land (~1019 ha) overall, while bush berries (~0.2 ha) occupy the least. The total acreage is provided in Table 1 and Fig 2.
Over 2019–2023, TC1 and TC2 show the strongest link between NDVI and ETa, while TC6 (in the first half of the WY, Oct–Mar) and TC5 (in the second half of the WY, Apr–Sep) are the weakest. In the first half of the water year, correlations range from very weak in TC6 to strong in TC2 and NDVI explains only a modest share of ETa variability (roughly the lower–mid tens of percent); by the second half, correlations strengthen, peaking in TC1–TC2, and NDVI explains a larger share of ETa (roughly one-third to about two-thirds). All-parcel regressions agree, with correlations of about 0.40–0.66 in the first half and 0.58–0.80 in the last half of the WY (Fig 4A and Fig 4B). Overall, NDVI tracks consumptive use much better in the second half of the WY than in the first. Based on the crop distribution and consumptive use patterns, it was evident that the truck crops are diverse, high-value, often multi-cropped, and comprise the largest area (~3859 ha) than vineyard and hay crops, exhibiting strong interannual, seasonal, and sub-seasonal variation in consumptive use (Fig 4 and 5).
3.1.2. Vineyards.
Four clusters emerged within the vineyard category from 707 parcels in the SCV study region. Significant differences in NDVI signal and consumptive use were observed across the four clusters (Fig 8 and S2 Appendix). The mean annual consumptive-use CV for vineyards was 11%, ranging from 9% in 2019 to 14% in 2021 (Fig 6). All 707 vineyard parcels are wine grapes, and both clustering methods divide them into four area-ranked clusters. Under only-NDVI, ETa cluster means cover 2,701–3,363 mm (1.2-fold), with within-cluster CVs of 18–26%; this CV range is comparable to the crop-type CV of 24%, the highest among the three crops. NDVI+ETa widens the cluster-mean range to 2,078-3,445 mm (1.7-fold) and tightens three of the four within-cluster CVs to 14–16%. At VC1, the two methods emphasize complementary features. The only-NDVI VC1 cluster lands at moderate ETa (2,782 mm) and identifies stable, mid-intensity vineyards with typical canopy phenology. The NDVI+ETa VC1 cluster lands at the high end (3,445 mm) and identifies the highest-volume water users, consistent with fully irrigated mature vineyards on alluvial soils. Together, the two VC1 definitions outline complementary dominant types: one defined by typical phenology and the other by typical consumptive-use volume. Within the NDVI+ETa scheme, VC1 carries the highest mean NDVI (0.369) and the lowest interannual variability (water-year CV = 5.8%). VC2 and VC4 have near-identical mean ETa but peak in different months (March versus May). VC3 is the outlier: it pairs high mean NDVI (0.346) with the lowest ETa, the steepest declining trend (-23 mm/yr), and the highest CV (32%). Both methods independently flag a high-variability tail subgroup, only-NDVI VC4 at CV = 26% and NDVI+ETa VC3 at CV = 32%. The labels differ in rank, but they point to a similar set of difficult-to-cluster parcels.
Vineyard clusters occupy the smallest total cropped area (~637 ha) relative to truck crops and hay crops (Table 2). Text B, Text C, and Table D to Table I in S2 Appendix, reports the corresponding within-cluster and between-cluster statistical differences.
The VC1, VC2, and VC4 clusters generally track precipitation, whereas VC3, with the largest ETa but lower and later-peaking NDVI, is more ETa-driven. In the first half of the WY (Oct-Mar), NDVI-ETa correlations are weak to moderate (roughly 0.3-0.5 across clusters, strongest in VC2), and regressions suggest that NDVI explains only a small share of ETa variability (Fig 7A). In the second half (Apr-Sep), when vine growth and irrigation peak, correlations strengthen (about 0.4-0.6, highest in VC3), and regression fits indicate that NDVI accounts for roughly one-quarter to one-third of ETa variation across vineyard clusters (Fig 7B).Overall, the vineyard consumptive use was observed to be moderately decreased from 2021 to 2023 (see Fig 6 CVs), consistent with improved irrigation control and less interannual volatility. These patterns (especially VC3) align well with cover-cropped and high vigor blocks (e.g., VC1/VC2/VC4) that show precipitation correlation consistent with stronger topographic soils and winter cover, as shown in Fig 8. The mild downward trend in vineyard consumptive use (like VC2) plausibly aligns with deficit irrigation adoption to sustain quality, widely reported in the literature and locally observed.
3.1.3. Hay crops.
The hay crop category produced three clusters (HC) from 438 parcels in the SCV. This crop group includes grain, pasture, and field crops (Fig 9). Annual CV averaged 12%, ranging from about 5% in 2019 to about 16% in 2021 (Table 3 and Fig 10). Hay crops occupied the second-largest cultivated area, roughly 2,987 ha, after truck crops. Hay produced three area-ranked clusters under either method, with closer cross-method agreement than truck crops or vineyards. Under only-NDVI, cluster means of ETa cover 2,836-3,491 mm (1.2-fold), and within-cluster CVs range from 10% to 22%. Under NDVI+ETa, cluster means span a slightly wider 2,723-3,684 mm range (1.4-fold), while within-cluster CVs tighten to 13–15%. The methods agree at HC1: both identify a low-ETa, rainfall-supplemented winter/spring grass group as the dominant cluster. They also agree at HC3, where both isolate a tight, high-ETa perennial-forage group. The only-NDVI HC3 cluster reaches the lowest CV among any only-NDVI cluster (10%). The intermediate HC2 cluster illustrates the added value of combining the two views. Under only-NDVI, HC2 is diffuse (CV = 22%) and overlaps HC1 in mean ETa. Under NDVI+ETa, HC2 tightens to CV = 13%, aligns with a warm-season annual signature that peaks in August, and separates clearly from the dominant HC1. Table 3 and Text B, Text C, and Table D to Table I in S2 Appendix, provide the corresponding cluster statistics.
The inset plot is the coefficient of variation.
HC1 and HC2 group mostly single-season forage/pasture fields, while HC3 captures the more intensive grain systems dominated by corn–sorghum–sudan plus some sunflowers, beans, and other grains. Cluster areas range from HC3 to HC1 (~650–1,400 ha), and on the crop side from dominant miscellaneous grains/hay and mixed pastures to tiny classes such as turf, wheat, and miscellaneous fields (~0.5–1,867 ha), as shown in Fig 9.
HC1 and HC2 have similar NDVI patterns, but HC1 uses less water, while HC3 reflects the grain-based rotation. In the first half of the WY (Oct–Mar), NDVI–ETa links are moderate–strong across clusters (correlations ~0.6–0.8), with NDVI explaining roughly 35–50% of ETa variability (Fig 11A). In the second half (Apr–Sep), correlations strengthen (~0.7–0.8; parcel-level ~0.75–0.85), so NDVI explains more than half the variation in ETa (about 55–75%), reflecting higher irrigation demand (Fig 11B). The HC3 shows more synchronized NDVI and ETa peaks, whereas HC1 and HC2 do not. This suggests that planting/rotation timing is likely part of the signal, but not the only explanation; atmospheric demand, irrigation timing, and management also likely contribute to the phase difference.
The clustering analysis revealed crop diversity and management contrasts within the hay/forage system. Consumptive-use variation is consistent with seasonal rotations, fallowing, varied irrigation technologies, clear within-season contrasts, and moderate interannual variation, particularly for HC3 in Fig 12.
Across truck crops, vineyards, and hay crops, total ETa magnitudes overlap substantially among crop categories (between-crop CV = 7.7% versus average within-crop CV = 20.1%; Table D in S2 Appendix). The management-relevant signal is therefore not simply that the three crop types use different amounts of water. Instead, within each crop type, parcels separate into distinct consumptive-use groups that differ in seasonal concentration of water use (warm-season ETa share, Truck/Vine/Hay = 70.9%, 69.6%, 67.7%), interannual variability (water-year CV = 8.7%, 8.1%, 12.1%), and canopy greenness (CV of mean NDVI = 33.3%, 12.7%, 20.2%; Table E in S2 Appendix). The only-NDVI clustering partitions each crop type into phenological subgroups (six for truck crops, four for vineyards, and three for hay crops), capturing when canopies green up and senesce. The NDVI+ETa partition adds ETa magnitude and interannual trend dimensions that scale-invariant NDVI-shape methods cannot resolve. Together, the two approaches show that within-crop variability exceeds between-crop differences across ETa magnitude, phenology, greenness, and interannual responsiveness, supporting crop-specific stratified clustering rather than treating crop type alone as a sufficient predictor of water demand (S1 Appendix and Table D to Table F in S2 Appendix).
4. Discussion
This study addresses a central gap in monitoring and managing agricultural consumptive use across diverse cropping systems in the Santa Clara Valley. Valley Water’s current Table of Averages approach assumes relatively uniform water use within broad crop categories and therefore obscures management-relevant parcel-scale variation. The results show that truck crops, vineyards, and hay crops differ in timing and management context, but they also show that heterogeneity within each crop category is larger than the difference in total ETa among crop categories. This within-crop variability is the main reason clustering is useful for monitoring and management.
4.1. Clustering approach versus traditional averaging
The clustering results show that broad labels such as truck crops, vineyards, and hay crops conceal substantial variation in phenology, multi-cropping, and irrigation practices. This finding is consistent with earlier research showing that advanced irrigation technologies, cover crops, and mixed cropping can reduce the accuracy of conventional ETa estimates based on broad crop categories [51–54].
Allowing phenology to influence group formation through NDVI time series captured differences in crop sequencing, cover cropping, spatial differentiation, and fallow periods that would be hidden by a single mean curve [55,56]. Cross-validation supports phenology as a useful predictor of consumptive use: NDVI-guided clusters reduced parcel-scale ETa error for truck crops by about one-third relative to a single profile [32,57]. The two-track design should therefore be read as a practical sequence rather than a competition between methods. Only-NDVI offers a transferable, ET-independent first pass for regions where reliable ETa data are limited, while NDVI+ETa provides stronger consumptive-use separation where ETa products are available and locally credible. Simpler ETa-only summaries may still be sufficient for some management applications, and additional clusters are not automatically better when marginal predictive gains are small. The selected k values therefore balance interpretability, cohesion/separation, predictive performance, and operational usefulness. Table F in S2 Appendix, shows that NDVI+ETa increases between-cluster separation while reducing average within-cluster CV relative to only-NDVI for the same crop categories.
4.2. Heterogeneity and agility in truck crops
Truck crops were expected to show the greatest within-season and interannual variability because they are high-value, short-season, and often double-cropped. The clustered patterns support this expectation and align with local knowledge of intensive rotations and rapid crop turnover. Compared with more narrowly planted vegetable clusters, multi-cropped vegetable clusters show stronger NDVI-ETa coupling during the main irrigation season and larger seasonal swings, reflecting sequences such as leafy greens rotated with solanaceous crops. These trends are consistent with local practice, where precision fertigation influences technology selection, drip and micro-sprinklers predominate, and sprinklers are retained primarily for germination [58,59]. From a programmatic perspective, the truck clusters identify multi-cropping vegetable systems as strong candidates for irrigation tuning and in-season scheduling support, because small adjustments can affect a large share of consumptive use.
These patterns are consistent with Santa Clara Valley truck-crop systems: short rotations, multi-cropping, and parcel-level management differences create distinct water-use signatures within the same broad crop label. The Sankey diagram links each NDVI+ETa cluster to a crop type and a management style. TC2 and TC4 both include peppers, lettuce, and miscellaneous truck crops, yet they occupy different phenological niches. TC2 captures spring-peaking, likely double-cropped parcels that rotate cool- and warm-season vegetables. TC4 covers warm-season-dominant operations on bare or mulched ground. This pair illustrates one water-user archetype: similar consumptive-use totals achieved through different management. TC1’s August peak matches a warm-season row-crop signature on valley-floor clay loams. Its rising ETa trend probably reflects increased atmospheric demand from higher vapor pressure deficit [60,61] or more intensive cropping. TC6’s persistent canopy and high consumptive use point to extended-season production, possibly with greenhouses or high tunnels. The two low-ETa clusters tell different stories. TC3 likely captures parcels easing out of production along the urban-agricultural fringe. TC5’s substantial consumptive use without a visible canopy may indicate greenhouses or bare-soil evaporation from irrigated, unplanted fields. TC1 versus TC5 shows a second archetype: different consumptive use through different management. TC1 versus TC6 shows a third: different consumptive use within similar broad management, with both representing warm-season vegetable systems at different intensities. The two clustering methods complement each other across these archetypes. Only-NDVI identifies when canopies green up and senesce. NDVI+ETa identifies consumptive-use magnitude and trend. Together they reinforce the six-group typology. Agreement on TC1 anchors the dominant water-user group, while NDVI+ETa resolution of TC3, TC5, and TC6 surfaces smaller subgroups that only-NDVI cannot separate. LandIQ statewide crop mapping [29] is more accurate than federal alternatives [30], but it applies a single-year crop label across the multi-year window. Rotations, fallowing transitions, and conversions during the study period therefore remain unobserved.
4.3. Spatial differentiation and stability in vineyards
Because vineyards are perennial systems with long-term investment, strong spatial differentiation was expected from topography, soil, vine age, and quality goals, with more muted temporal shifts [62–66]. The clustered vineyard modes are consistent with that expectation and with documented use of cover crops and selective deficit irrigation; they range from vigorous, later-peaking canopies to more conservative, precipitation-tracking systems [66,67]. Overall vineyard consumptive use declines slightly during the study period, while NDVI-ETa relationships strengthen in the second half of the WY when vine water demand and irrigation peak. Taken together, these patterns are consistent with improved scheduling and deficit strategies in SCV vineyards [25], and remotely sensed ETa/NDVI can help refine decisions such as deficit-irrigation timing and cover-crop termination.
Vineyards in this valley appear to be separated more by management intensity and vineyard condition than by grape variety. VC1 represents fully irrigated mature vineyards, possibly on deep valley-floor alluvial soils where vine vigor is not deliberately limited. These parcels require the most water per hectare and offer the greatest potential savings if regulated deficit irrigation (RDI) is adopted. California field trials show RDI can reduce seasonal applied water by 20–40%, with modest yield reductions often offset by improved wine quality [68,69]. The VC2/VC4 pair illustrates similar consumptive use achieved through different management. At near-identical ETa, VC2’s higher NDVI and late-winter peak suggest mature vineyards with active inter-row cover cropping [70,71]. VC4’s suppressed canopy and May peak suggest heavily pruned, low-vigor vines on shallower foothill soils, managed for fruit quality through deliberate canopy restriction. The VC1/VC3 contrast illustrates the converse: different consumptive use within similar broad management. Both are wine-grape parcels under similar biological constraints, but VC1 appears irrigated to full vine vigor, while VC3 likely aggregates young vineyards still establishing roots, mature vineyards under severe RDI, and parcels transitioning out of viticulture near the urban fringe. The two methods reinforce this interpretation. Because vineyards share a common biological phenology, only-NDVI has limited signal to distinguish water-use magnitude. Its compressed cluster-mean range (1.2-fold) captures canopy-rhythm differences more than absolute ETa differences. NDVI+ETa adds the management-intensity and trend signal that distinguishes wine-grape parcels in practice, widening the cluster-mean range to 1.7-fold. The shared identification of a high-CV tail subgroup, only-NDVI VC4 at 26% and NDVI+ETa VC3 at 32%, indicates that some heterogeneity is structural rather than methodological. Vine age, planting year, and parcel-development status are missing features, and no remote-sensing time series, whether phenology- or magnitude-based, can fully supply them. LandIQ’s static single-year crop map [29,30] compounds this limitation because parcels undergoing vineyard establishment, removal, or conversion carry a fixed crop identity that may not match their actual status across the analysis window.
4.4. Within-season contrasts in hay and forage systems
With significant confusion in traditional mapping between forage, field crops, and related cover crops, it was anticipated that hay and forage would exhibit strong within-season contrasts between pasture-like uses and grain/forage rotations [72–74]. Grain/forage rotations (such as corn–sorghum–Sudan), which show higher late-season ETa and unique phenological signatures, are clearly distinguished from pasture-dominated systems by the clusters. It is plausible that a significant proportion of the management signals—such as stocking intensity, cutting frequency, seasonal fallowing (e.g., HC3), and rotation timing—that are otherwise challenging to observe may be inferred through phenology. This inference is supported by the robustness of NDVI–ETa relationships across hay clusters, particularly during the primary irrigation season. These categories are particularly helpful for focusing on soil-moisture monitoring and rotational planning because hay systems make up a significant portion of total consumptive use.
The Sankey diagram maps each cluster onto a forage system. HC2’s August NDVI peak and high seasonal cycling match warm-season annuals such as corn silage and sorghum/sudan grass. These crops are planted in late spring, fully canopied by August, and bare after fall harvest, concentrating water demand into a four-month window. HC3’s year-round greenness and spring NDVI peak are consistent with alfalfa or intensively managed perennial pasture. Alfalfa sustains six to eight cuttings per year under California’s Mediterranean climate [75] and is the heaviest consumptive user among common forage crops, at roughly 1.1-1.5 m annually statewide (equivalent to 3.5-5.0 acre-feet) [76,77]. HC2 and HC3 illustrate similar annual consumptive use achieved through different management. Their total ETa converges in NDVI+ETa, but HC2 needs high-volume seasonal delivery while HC3 needs moderate continuous supply, a distinction that matters for water-delivery scheduling. HC1 maps mostly to miscellaneous grains/hay and mixed pastures, where irrigation supplements rather than replaces rainfall. Its high interannual variability and low warm-season share reflect dependence on winter/spring precipitation. On clay-dominant valley floors, stored winter moisture can sustain grass growth into early summer. That reduces irrigation in wet years but creates sharp demand spikes in drought years, a pattern expected to intensify under projected increases in California precipitation volatility [78]. Adaptive water allocations therefore make more sense than fixed allocations for this dominant parcel class. Hay shows the closest agreement of the three crops between the two methods. Each forage subtype carries a distinct phenological signature, so only-NDVI clustering aligns well with the underlying water-user grouping. NDVI+ETa adds quantitative detail by resolving consumptive-use magnitude and interannual responsiveness: HC2’s seasonal concentration, HC3’s near-perennial demand, and HC1’s climate-driven year-to-year swings. The two methods identify the same three water-user groups from complementary perspectives. Only-NDVI identifies timing; NDVI+ETa identifies magnitude and trend. This convergence supports the three-group typology as a basis for crop-aware water management.
4.5. Implications for monitoring and management
At the district scale, the study demonstrates that static averages and district-scale mass balance are insufficient to identify which agricultural parcels drive variability in consumptive use, even where overall irrigation efficiency appears high [25]. Extension and incentive programs can use the resolved consumptive-use groups within each crop type because they are internally coherent and aligned with recognizable on-farm systems, including multi-cropped vegetables, vigor-differentiated vineyards, and pasture versus grain/forage systems.
According to Rallings et al. [25], these groups also highlight limitations of existing field-based monitoring, where many operations lack routine measurements and both technology and climate are changing quickly. High-resolution ETa products from platforms such as OpenET, combined with NDVI, ML, and ground observations, provide a path toward more dynamic and transparent accounting of agricultural consumptive use [32,59,79,80].
Overall, the findings support the central hypothesis that remotely sensed ETa and NDVI, analyzed with clustering and pattern-mining techniques [81], can move water accounting beyond coarse tables of averages and toward a more management-relevant view of how water is used across agricultural systems. This is important not only for improving irrigation efficiency and conservation outcomes, but also for understanding how water availability and water distribution across crops and technologies affect agricultural production and community well-being in a changing climate.
5. Limitations
This research has several limitations that may affect the validity and applicability of the conclusions and recommendations. First, remote sensing data provide high-resolution spatial and temporal insights but remain susceptible to errors from cloud cover, sensor quality, and processing choices. In addition, NDVI-based clustering alone cannot fully represent the complexity of consumptive use in mixed cropping systems, where many variables influence water demand [13,82,83]. The study also did not validate parcel-scale consumptive use against ground-based measurements and instead relied on satellite-based OpenET estimates.
Second, a single-year crop map was used as a fixed spatial reference for extracting multi-year NDVI and ETa data. This choice provides a pragmatic regional sampling frame, supported by previous evidence that SCV crop maps were relatively stable over the last decade [25]. Nevertheless, crop rotations, parcel turnover, and annual land-use changes remain important limitations for interpreting multi-year parcel-level results.
Third, the study did not directly examine field-level planting and harvest records, temperature or climate-stress responses, cover cropping, fallowing, soil properties and topography, irrigation technology and scheduling, crop age, vineyard condition, or conservation practices. The analysis instead focused on major crop groups, growth phenology, and satellite-derived consumptive-use data. As a result, the specific SCV cluster patterns are not directly transferable to locations with different crop mixes, management practices, soils, climate, ETa-product reliability, or water-delivery institutions without local validation. In such settings, the only-NDVI workflow can provide a transferable first-pass phenology classification, but ETa-based interpretation should be added only after the reliability of local ET data has been assessed.
Fourth, the study period was limited to water years 2019–2023, which is too short to capture long-term hydroclimatic variability. This window includes the 2020–2022 drought and the very wet 2023 water year, but additional wet, dry, and near-normal years would be needed to distinguish persistent management patterns from short-term climate responses.
6. Recommendations for future work
Several future research and implementation steps could strengthen agricultural consumptive-use monitoring in the SCV. The recommendations below are organized by likely stakeholder responsibility and approximate ease of implementation.
- Valley Water and district water managers should use the identified clusters to design stratified monitoring, sampling, and outreach programs that target the parcels most likely to drive consumptive-use variability.
- Growers, extension advisors, and irrigation consultants should pair cluster information with field-level irrigation scheduling, soil-moisture monitoring, and technology assessments to evaluate how irrigation methods and conservation practices affect consumptive-use efficiency.
- Researchers and agency partners should integrate higher-resolution monitoring data, including field observations, Planet-scale imagery, EC towers, or other ground-based measurements, to validate and refine satellite-derived cluster assignments.
- Policy makers, Valley Water staff, growers, and community stakeholders should co-develop decision tools, incentives, and training materials that translate cluster-based water accounting into practical conservation and allocation strategies.
- Researchers and water managers should explore new datasets on irrigation scheduling, regulatory compliance, crop transitions, and grower behavior so that geospatial AI and spatial analytics can better explain complex water-use patterns and future trends.
7. Conclusions
This study demonstrates that parcel-scale remote sensing and unsupervised clustering can convert broad agricultural crop labels into operational consumptive-use groups for water accounting. In the Santa Clara Valley, truck crops, vineyards, and hay crops contain substantial within-crop heterogeneity that is obscured by conventional crop averages. Only-NDVI clustering provides a scalable, ET-independent phenology-first view of crop timing, which is useful for transfer to regions where reliable ETa data may be unavailable or delayed. NDVI+ETa clustering adds consumptive-use magnitude and interannual trend where credible ETa estimates are available, improving separation among water-user groups. The resulting clusters identify management-relevant patterns in multi-cropped vegetables, vigor-differentiated vineyards, and pasture versus grain/forage systems, while also showing that direct attribution requires field, metering, or management records. Cluster-based monitoring can therefore help water managers target sampling, conservation incentives, and adaptive allocation in groundwater-dependent agricultural regions, provided that local crop maps, climate conditions, management practices, and ET-data reliability are validated before transfer to other settings.
Supporting information
S1 Appendix. The descriptive statistics of the agricultural consumptive use (~ETa) for each major crop category.
This shows a wide range of variations that are obscured by the averages and therefore form the basis for this study.
https://doi.org/10.1371/journal.pwat.0000416.s001
(DOCX)
S2 Appendix. The number of cluster possibilities in each major crop category and the statistics of the analysis were used to finalize the number of clusters.
Detailed analysis of within-cluster and between-cluster NDVI and ETa. Also include the per-cluster statistics and agricultural consumptive use statistics for all three major crop types.
https://doi.org/10.1371/journal.pwat.0000416.s002
(DOCX)
Acknowledgments
We thank the contributions of A. Rallings, F. Gamiño, N. Santos, K. Moyers, and A. Silverstein in the earlier portion of this study. We acknowledge the peer reviewers for their insightful and constructive comments that helped shape the manuscript.s
References
- 1. Rosa L. Adapting agriculture to climate change via sustainable irrigation: biophysical potentials and feedbacks. Environ Res Lett. 2022;17(6):063008.
- 2. Döll P. Impact of Climate Change and Variability on Irrigation Requirements: A Global Perspective. Climatic Change. 2002;54(3):269–93.
- 3. Qin Y, Abatzoglou JT, Siebert S, Huning LS, AghaKouchak A, Mankin JS. Agricultural risks from changing snowmelt. Nature Climate Change. 2020;10(5):459–65.
- 4. Scanlon BR, Faunt CC, Longuevergne L, Reedy RC, Alley WM, McGuire VL, et al. Groundwater depletion and sustainability of irrigation in the US High Plains and Central Valley. Proc Natl Acad Sci U S A. 2012;109(24):9320–5. pmid:22645352
- 5. Assouline S, Russo D, Silber A, Or D. Balancing water scarcity and quality for sustainable irrigated agriculture. Water Resources Research. 2015;51(5):3419–36.
- 6. Nolz R. A review on the quantification of soil water balance components as a basis for agricultural water management with a focus on weighing lysimeters and soil water sensors. Journal of Land Management, Food and Environment. 2016;67(3):133–44.
- 7. Moyers K, Sabie R, Waring E, Preciado J, Naughton CC, Harmon T, et al. A Decade of Data‐Driven Water Budgets: Synthesis and Bibliometric Review. Water Resources Research. 2023;59(11).
- 8.
Andales AA, Chávez JL, Bauder TA. Irrigation scheduling: The water-balance approach. Fort Collins, CO: Colorado State University Extension. 2011. https://itc.tamu.edu/files/2018/05/04707.pdf
- 9. Evans RG, Sadler EJ. Methods and technologies to improve efficiency of water use. Water Resources Research. 2008;44(7).
- 10. Anapalli SS, Ahuja LR, Gowda PH, Ma L, Marek G, Evett SR, et al. Simulation of crop evapotranspiration and crop coefficients with data in weighing lysimeters. Agricultural Water Management. 2016;177:274–83.
- 11. Moyers K, Abatzoglou JT, Escriva-Bou A, Medellín-Azuara J, Viers JH. An invisible water surcharge: Climate warming increases crop water demand in the San Joaquin Valley’s groundwater-dependent irrigated agriculture. PLOS Water. 2024;3(3):e0000184.
- 12. Levidow L, Zaccaria D, Maia R, Vivas E, Todorovic M, Scardigno A. Improving water-efficient irrigation: Prospects and difficulties of innovative practices. Agricultural Water Management. 2014;146:84–94.
- 13. Brar AS, Kaur K, Sindhu VK, Tsolakis N, Srai JS. Sustainable water use through multiple cropping systems and precision irrigation. Journal of Cleaner Production. 2022;333:130117.
- 14. André FJ, Buendía A, Santos-Arteaga FJ. Efficient water use and reusing processes across Spanish regions: A circular data envelopment analysis with undesirable inputs. Journal of Cleaner Production. 2024;434:139929.
- 15. Kalyanaraman A, Burnett M, Fern A, Khot L, Viers J. Special report: The AgAID AI institute for transforming workforce and decision support in agriculture. Computers and Electronics in Agriculture. 2022;197:106944.
- 16. Yang L, Driscol J, Sarigai S, Wu Q, Chen H, Lippitt CD. Google Earth Engine and Artificial Intelligence (AI): A Comprehensive Review. Remote Sensing. 2022;14(14):3253.
- 17. Khosla R, Westfall DG, Reich RM, Mahal JS, Gangloff WJ. Spatial Variation and Site-Specific Management Zones. Geostatistical Applications for Precision Agriculture. Springer Netherlands. 2010. p. 195–219.
- 18. Gómez C, White JC, Wulder MA. Optical remotely sensed time series data for land cover classification: A review. ISPRS Journal of Photogrammetry and Remote Sensing. 2016;116:55–72.
- 19. Zhang L, Weng Q, Shao Z. An evaluation of monthly impervious surface dynamics by fusing Landsat and MODIS time series in the Pearl River Delta, China, from 2000 to 2015. Remote Sensing of Environment. 2017;201:99–114.
- 20. Balafoutis A, Koundouras S, Anastasiou E, Fountas S, Arvanitis K. Life Cycle Assessment of Two Vineyards after the Application of Precision Viticulture Techniques: A Case Study. Sustainability. 2017;9(11):1997.
- 21. Balafoutis A, Beck B, Fountas S, Vangeyte J, Wal T V der, Soto I. Precision agriculture technologies positively contributing to GHG emissions mitigation, farm productivity and economics. Sustainability. 2017;9(8):8.
- 22.
County of Santa Clara, Crop Report. California Department of Food & Agriculture. 2022. https://files.santaclaracounty.gov/2023-10/cropreport2023final.pdf?VersionId=fEZGTRIBNm.o8V7uly8PqIkvIl035bH2
- 23. Chaussard E, Milillo P, Bürgmann R, Perissin D, Fielding EJ, Baker B. Remote Sensing of Ground Deformation for Monitoring Groundwater Management Practices: Application to the Santa Clara Valley During the 2012–2015 California Drought. JGR Solid Earth. 2017;122(10):8566–82.
- 24.
ERA Economics. Santa Clara Valley Water District Crop Irrigation Water Use Update. ERA Economics. 2014. https://www.valleywater.org/sites/default/files/2018-01/Crop%20Water%20Use%20final%20report.pdf
- 25.
Rallings AM, Santos N, Gamiño F, Medellín-Azuara J, Viers JH. Santa Clara Valley Water District Agricultural Water Use Baseline Study. University of California, Merced. 2022.
- 26.
County of Santa Clara, Santa Clara Valley Open Space Authority. Santa Clara Valley Agriculture Plan. County of Santa Clara and the Santa Clara Valley Open Space Authority. 2018. https://stgenpln.blob.core.windows.net/document/SCV_ActionPlan.pdf
- 27. Spurlin SM, Lambrecht SC. Rapid responses to drought in a rare California annual (San Francisco collinsia, Collinsia multicolor). Plant Ecol. 2024;225(9):997–1007.
- 28. Medellín-Azuara J, Escriva-Bou A, Rodríguez-Flores JM, Cole SA, Abatzoglou JT, Viers JH. Economic impacts of the 2020–22 drought on California agriculture. 2022. https://cawaterlibrary.net/wp-content/uploads/2022/11/20AmSf-Economic_Impact_CA_Drought_V01.pdf
- 29. Land IQ. https://www.landiq.com/land-use-mapping 2023. 2024 June 5.
- 30. Espinoza V, Booth LA, Viers JH. Land use misclassification results in water use, economic value, and GHG emission discrepancies in California’s high-intensity agriculture region. Sustainability. 2023;15(8):8.
- 31. Heilmann M. Maxar Blog. https://blog.maxar.com/earth-intelligence/2021/land-iq-leverages-securewatch-to-determine-land-use-change-in-california 2021. 2024 June 5.
- 32. Melton FS, Huntington J, Grimm R, Herring J, Hall M, Rollison D, et al. OpenET: Filling a Critical Data Gap in Water Management for the Western United States. J American Water Resour Assoc. 2021;58(6):971–94.
- 33. Luzio MD, Johnson GL, Daly C, Eischeid JK, Arnold JG. Constructing Retrospective Gridded Daily Precipitation and Temperature Datasets for the Conterminous United States. 2008.
- 34. Daly C, Bryant K. The PRISM Climate and Weather System – An Introduction. https://prism.oregonstate.edu/documents/PRISM_history_jun2013.pdf 2013.
- 35.
Louis J, Pflug B, Main-Knorn M, Debaecker V, Mueller-Wilm U, Iannone RQ, et al. Sentinel-2 Global Surface Reflectance Level-2a Product Generated with Sen2Cor. In: IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, 2019. 8522–5.https://doi.org/10.1109/igarss.2019.8898540
- 36. Farbo A, Sarvia F, De Petris S, Basile V, Borgogno-Mondino E. Forecasting corn NDVI through AI-based approaches using sentinel 2 image time series. ISPRS Journal of Photogrammetry and Remote Sensing. 2024;211:244–61.
- 37. Chakhar A, Hernández-López D, Ballesteros R, Moreno MA. Irrigation detection using Sentinel-1 and Sentinel-2 time series on fruit tree orchards. Remote Sensing. 2024;16(3):3.
- 38. Zhang Z, Tang P, Huo L, Zhou Z. MODIS NDVI time series clustering under dynamic time warping. Int J Wavelets Multiresolut Inf Process. 2014;12(05):1461011.
- 39. Gao F, Anderson M, Daughtry C, Karnieli A, Hively D, Kustas W. A within-season approach for detecting early growth stages in corn and soybean using high temporal and spatial resolution imagery. Remote Sensing of Environment. 2020;242:111752.
- 40.
Hautamaki V, Nykanen P, Franti P. Time-series clustering by approximate prototypes. In: 2008 19th International Conference on Pattern Recognition, 2008. 1–4. https://doi.org/10.1109/icpr.2008.4761105
- 41. Meesrikamolkul W, Niennattrakul V, Ratanamahatana CA. Shape-Based Clustering for Time Series Data. Lecture Notes in Computer Science. Springer Berlin Heidelberg. 2012. p. 530–41.
- 42. Möller-Levet CS, Klawonn F, Cho K-H, Wolkenhauer O. Fuzzy Clustering of Short Time-Series and Unevenly Distributed Sampling Points. Lecture Notes in Computer Science. Springer Berlin Heidelberg. 2003. p. 330–40.
- 43.
Kalpakis K, Gada D, Puttagunta V. Distance measures for effective clustering of ARIMA time-series. In: Proceedings 2001 IEEE International Conference on Data Mining. 273–80. https://doi.org/10.1109/icdm.2001.989529
- 44. D’Urso P, Cappelli C, Di Lallo D, Massari R. Clustering of financial time series. Physica A: Statistical Mechanics and its Applications. 2013;392(9):2114–29.
- 45. Ohana-Levi N, Gao F, Knipper K, Kustas WP, Anderson MC, del Mar Alsina M. Time-series clustering of remote sensing retrievals for defining management zones in a vineyard. Irrig Sci. 2022;40(4):801–15.
- 46. Filintas A, Gougoulias N, Kourgialas N, Hatzichristou E. Management soil zones, irrigation, and fertigation effects on yield and oil content of Coriandrum sativum L. using precision agriculture with fuzzy k-means clustering. Sustainability. 2023;15(18):13524.
- 47. González-Dugo MP, Escuin S, Cano F, Cifuentes V, Padilla FLM, Tirado JL, et al. Monitoring evapotranspiration of irrigated crops using crop coefficients derived from time series of satellite images. II. Application on basin scale. Agricultural Water Management. 2013;125:92–104.
- 48. Ali M, Alqahtani A, Jones MW, Xie X. Clustering and Classification for Time Series Data in Visual Analytics: A Survey. IEEE Access. 2019;7:181314–38.
- 49. Nirmal S. Comparative study between k-means and k-medoids clustering algorithms. Int Res J Eng Technol. 2019;839(1):839–44.
- 50. Caliński T, Harabasz J. A dendrite method for cluster analysis. Communications in Statistics. 1974;3(1):1–27.
- 51. Falalakis G, Gemitzi A. A simple method for water balance estimation based on the empirical method and remotely sensed evapotranspiration estimates. Journal of Hydroinformatics. 2020;22(2):440–51.
- 52. Jovanovic N, Pereira LS, Paredes P, Pôças I, Cantore V, Todorovic M. A review of strategies, methods and technologies to reduce non-beneficial consumptive water use on farms considering the FAO56 methods. Agricultural Water Management. 2020;239:106267.
- 53. Nikolaou G, Neocleous D, Christou A, Kitta E, Katsoulas N. Implementing Sustainable Irrigation in Water-Scarce Regions under the Impact of Climate Change. Agronomy. 2020;10(8):1120.
- 54. Zinkernagel J, Maestre-Valero JF, Seresti SY, Intrigliolo DS. New technologies and practical approaches to improve irrigation management of open field vegetable crops. Agricultural Water Management. 2020;242:106404.
- 55. Schauer M, Senay GB. Characterizing Crop Water Use Dynamics in the Central Valley of California Using Landsat-Derived Evapotranspiration. Remote Sensing. 2019;11(15):1782.
- 56. Torres N, Yu R, Martínez-Lüscher J, Kostaki E, Kurtural SK. Effects of Irrigation at Different Fractions of Crop Evapotranspiration on Water Productivity and Flavonoid Composition of Cabernet Sauvignon Grapevine. Front Plant Sci. 2021;12:712622. pmid:34539704
- 57. Volk JM, Huntington JL, Melton FS, Allen R, Anderson M, Fisher JB, et al. Assessing the accuracy of OpenET satellite-based evapotranspiration data to support water resource and land management applications. Nat Water. 2024;2(2):193–205.
- 58. Angelakιs AN, Zaccaria D, Krasilnikoff J, Salgot M, Bazza M, Roccaro P, et al. Irrigation of World Agricultural Lands: Evolution through the Millennia. Water. 2020;12(5):1285.
- 59. Sun AY, Scanlon BR. How can big data and machine learning benefit environment and water management: a survey of methods, applications, and future directions. Environ Res Lett. 2019;14(7):073001.
- 60. Yuan W, Zheng Y, Piao S, Ciais P, Lombardozzi D, Wang Y. Increased atmospheric vapor pressure deficit reduces global vegetation growth. Science Advances. 2019;5(8):eaax1396.
- 61. Novick KA, Ficklin DL, Grossiord C, Konings AG, Martínez-Vilalta J, Sadok W, et al. The impacts of rising vapour pressure deficit in natural and managed ecosystems. Plant Cell Environ. 2024;47(9):3561–89. pmid:38348610
- 62. Bramley RGV. Understanding variability in winegrape production systems 2. Within vineyard variation in quality over several vintages. Aust J Grape Wine Res. 2005;11(1):33–42.
- 63. Bramley RGV, Hamilton RP. Understanding variability in winegrape production systems. Australian Journal of Grape and Wine Research. 2008;10(1):32–45.
- 64. Centinari M, Filippetti I, Bauerle T, Allegro G, Valentini G, Poni S. Cover Crop Water Use in Relation to Vineyard Floor Management Practices. Am J Enol Vitic. 2013;64(4):522–6.
- 65. Mania E, Petrella F, Giovannozzi M, Piazzi M, Wilson A, Guidoni S. Managing Vineyard Topography and Seasonal Variability to Improve Grape Quality and Vineyard Sustainability. Agronomy. 2021 Jun;11(6):1142.
- 66. Mirás-Avalos JM, Araujo ES. Optimization of vineyard water management: challenges, strategies, and perspectives. Water. 2021;13(6):6.
- 67. Romero P, Navarro JM, Ordaz PB. Towards a sustainable viticulture: The combination of deficit irrigation strategies and agroecological practices in Mediterranean vineyards. A review and update. Agricultural Water Management. 2022;259:107216.
- 68. Fereres E, Soriano MA. Deficit irrigation for reducing agricultural water use. J Exp Bot. 2007;58(2):147–59.
- 69. Costello MJ, Patterson WK. Regulated Deficit Irrigation Effect on Yield and Wine Color of Cabernet Sauvignon in Central California. horts. 2012;47(10):1520–4.
- 70. Steenwerth K, Belina KM. Cover crops enhance soil organic matter, carbon dynamics and microbiological function in a vineyard agroecosystem. Applied Soil Ecology. 2008;40(2):359–69.
- 71. Ingels CA, Scow KM, Whisson DA, Drenovsky RE. Effects of cover crops on grapevines, yield, juice composition, soil microbial ecology, and gopher activity. Am J Enol Vitic. 2005;56(1):19–29.
- 72.
Havilah EJ. Forages and pastures | Annual forage and pasture crops – species and varieties. Elsevier. 2002. p. 1131–42.
- 73.
Moss N, Havilah EJ. Annual Forage and Pasture Crops – Establishment and Management. Academic Press. 2022. p. 711–31.
- 74.
Wortmann CS, Anderson BE, Redfearn DD. Perennial Grass or Grass-Legume in Rotation with Annual Crops. University of Nebraska-Lincoln, Extension. 2019.
- 75.
Summers CG, Putnam DH. Irrigated alfalfa management for Mediterranean and desert zones. UCANR Publications. 2008.
- 76. California Department of Water Resources. Statewide Agricultural Water Use Data 2016-2020 - California Natural Resources Agency. 2020. https://data.cnra.ca.gov/dataset/agricultural-water-use-data-2016-2020
- 77. Hanson B, Putnam D, Snyder R. Deficit irrigation of alfalfa as a strategy for providing water for water-short areas. Agricultural Water Management. 2007;93(1–2):73–80.
- 78. Swain DL, Langenbrunner B, Neelin JD, Hall A. Increasing precipitation volatility in twenty-first-century California. Nature Clim Change. 2018;8(5):427–33.
- 79. Cai W, Ullah S, Yan L, Lin Y. Remote Sensing of Ecosystem Water Use Efficiency: A Review of Direct and Indirect Estimation Methods. Remote Sensing. 2021;13(12):2393.
- 80. Huang R, Ma C, Ma J, Huangfu X, He Q. Machine learning in natural and engineered water systems. Water Res. 2021;205:117666. pmid:34560616
- 81. Thotli LR, Gugamsetty B, Kalluri ROR, Tandule CR, Kotalo RG, Akkiraju B, et al. Long‐term (2001–2020) trend analysis of temperature and rainfall and drought characteristics by in situ measurements at a tropical semi‐arid station from southern peninsular India. Intl Journal of Climatology. 2022;42(16):8928–49.
- 82. Jovanovic N, Pereira LS, Paredes P, Pôças I, Cantore V, Todorovic M. A review of strategies, methods and technologies to reduce non-beneficial consumptive water use on farms considering the FAO56 methods. Agricultural Water Management. 2020;239:106267.
- 83. Carlisle L, Esquivel K, Baur P, Ichikawa NF, Olimpi EM, Ory J, et al. Organic farmers face persistent barriers to adopting diversification practices in California’s Central Coast. Agroecology and Sustainable Food Systems. 2022;46(8):1145–72.