Satellites and Machine Learning Reveal Hidden Water Use Patterns in California Farms

In the fertile farmlands of California’s Santa Clara Valley, two fields growing the same crop can drink very different amounts of water. That hidden variability has long been invisible to water managers, who traditionally rely on broad crop categories and regional averages to estimate how much water agriculture actually consumes. A new study published in PLOS Water shows that a combination of satellite remote sensing and unsupervised machine learning can pull back that curtain, revealing striking parcel-by-parcel differences in water consumption that conventional accounting methods completely miss. The findings could reshape how groundwater-dependent regions plan for drought, allocate scarce supplies, and adapt to a changing climate.

The research team, led by Abid Sarwar and including Josué Medellín-Azuara, John T. Abatzoglou, and Joshua H. Viers, focused on a concept known as consumptive use. In water resources engineering, consumptive use refers to water that is taken up by plants and evaporated into the atmosphere and is therefore no longer available for downstream users or aquifer recharge. It is not the same as the water diverted from a canal or pumped from a well, because some of that diverted water returns to the system through runoff or deep percolation. Consumptive use is best approximated by actual evapotranspiration, the sum of evaporation from soil and transpiration through plant leaves, which is the quantity that ultimately determines how much water a farming operation truly removes from the landscape.

To measure evapotranspiration at the scale of individual fields, the researchers turned to OpenET, a publicly available platform that produces gridded estimates of actual evapotranspiration by combining multiple satellite-driven energy balance models into an ensemble product. They paired these evapotranspiration estimates with vegetation index data from the Sentinel-2 satellites, specifically the normalized difference vegetation index, or NDVI, which tracks the greenness and vigor of plant canopies through time. Because NDVI responds to canopy development, crop type, planting density, and stress, its seasonal trajectory carries a fingerprint of how each parcel is being managed. The team also incorporated precipitation data from the PRISM climate mapping system to place each parcel’s water use in its hydroclimatic context.

The study area covered 2,189 agricultural parcels spanning roughly 7,483 hectares of the Santa Clara Valley, an agriculturally productive and remarkably diverse region south of the San Francisco Bay. The analysis focused on three major crop categories: truck crops, which include vegetables and berries; vineyards; and hay crops. Data from 2019 through 2023 were assembled into time series for every parcel, creating a rich record of how vegetation greenness, water consumption, and rainfall varied seasonally and from year to year across the landscape.

Rather than assuming that all fields of a given crop behave alike, the researchers applied unsupervised machine learning, a family of algorithms that discovers natural groupings in data without any predefined labels. Two clustering approaches were tested: one using NDVI time series alone, and a second combining NDVI with the magnitude and trend of actual evapotranspiration. Both approaches identified 13 distinct consumptive-use clusters across the valley, but the composition was telling. Truck crops split into six clusters, vineyards into four, and hay crops into three. In other words, what appears on paper as a single crop category actually encompasses several operationally distinct water-use regimes.

The quantitative comparison between the two clustering strategies revealed why adding evapotranspiration information matters. The researchers evaluated cluster quality using the coefficient of variation, a normalized measure of dispersion, both within clusters and between cluster means. Under the NDVI-only approach, the ratio of between-cluster separation to average within-cluster variation was modest: 0.30 for truck crops, 0.46 for vineyards, and 0.74 for hay crops. When evapotranspiration magnitude and trend were added, those ratios jumped to 1.35, 1.06, and 1.07 respectively, meaning the clusters became far more distinct from one another relative to the scatter within them. Adding the water-use dimension also generally reduced the average within-cluster coefficient of variation. For context, the crop-wide coefficients of variation, the very heterogeneity that conventional averages conceal, stood at 17 percent for truck crops, 24 percent for vineyards, and 20 percent for hay crops.

Each crop category told its own story. Truck crops displayed the strongest seasonal and interannual heterogeneity of the three groups, which makes sense for a category encompassing vegetables and berries with different planting calendars, growth cycles, and irrigation demands. Vineyards showed clearer spatial differentiation and, notably, a modest decline in consumptive use over the study period, a pattern consistent with the widespread adoption of efficient irrigation technology and deliberate deficit irrigation strategies that premium winegrape growers use to control fruit quality. Hay crops separated into two fundamentally different systems: pasture-based operations and grain-forage systems, which have contrasting water delivery needs and different sensitivities to climate variability. A hay field that is actually irrigated pasture behaves very differently from one planted to a cereal forage crop, and the clustering algorithm detected that distinction without ever being told to look for it.

Crucially, the observed consumptive-use patterns aligned with known hydroclimatic stresses and management practices in the valley, including cover cropping, variation in irrigation technology, deficit irrigation, and seasonal fallowing. The authors are careful to note that directly attributing a specific cluster’s behavior to a specific practice would require independent evidence from field surveys, water permits, or metering data, which the remote sensing analysis alone cannot provide. Still, the correspondence between the statistical structure of the clusters and the agronomic realities on the ground lends credibility to the approach. The machine learning is not inventing patterns; it is surfacing real differences in how water moves through these agricultural systems.

The practical implications extend well beyond Santa Clara Valley. In regions where agriculture depends heavily on groundwater, regulators and irrigation districts need to know not just how much water a crop category consumes on average, but which individual parcels are consuming more or less than their peers and why. Parcel-scale consumptive-use clusters can support customized water-management plans, targeted irrigation monitoring, and conservation programs that reward growers who demonstrably reduce consumptive use rather than those who merely plant a low-use crop. During droughts, when fallowing decisions and curtailments are negotiated field by field, having an objective, remotely sensed baseline of who uses what, and how those patterns shift year to year, provides a common factual foundation for difficult allocation conversations.

The study also demonstrates a broader methodological lesson for the era of open environmental data. By fusing freely available Sentinel-2 imagery, the OpenET ensemble, and PRISM climate records with accessible machine learning tools, the researchers produced actionable water intelligence without deploying a single ground sensor across the study area. As satellite revisit times shorten and evapotranspiration products grow more accurate, similar analyses could be replicated in irrigated valleys worldwide, giving water managers in groundwater-dependent basins the fine-grained, adaptive picture of agricultural water use that legacy crop maps were never designed to deliver. For a state where every acre-foot counts, seeing water use field by field may prove as important as measuring it at the wellhead.

Subject of Research: Remote sensing and machine learning analysis of parcel-scale agricultural consumptive water use in California's Santa Clara Valley

Article Title: Identifying agricultural consumptive-use patterns to support adaptive water management in California’s Santa Clara Valley via remote sensing and machine learning

Article References: Sarwar, A., Medellín-Azuara, J., Abatzoglou, J. T., & Viers, J. H. (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

Image Credits: AI Generated

DOI: 10.1371/journal.pwat.0000416

Keywords: remote sensing, machine learning, evapotranspiration, consumptive use, groundwater management, Santa Clara Valley, Sentinel-2, OpenET, NDVI, agriculture, water management, California