Water demands of permitted and unpermitted cannabis cultivation in Northern California

The study area encompasses Humboldt and Mendocino counties in Northern California (figure 1). A representative sample constituting half of the watersheds (at the HUC-12 scale; USGS 2019) in the study area were randomly selected and systematically inspected for cannabis cultivation, following protocols established by Butsic and Brenner 2016. All cannabis farms in the selected watersheds were mapped from 2018 NAIP imagery by hand-digitizing polygon features of cultivated areas (Butsic et al 2018). The relative area of outdoor gardens and mixed-light growing infrastructure (greenhouses and hoop-houses) was used to calculate the proportion of outdoor cultivation per farm. Polygons of cannabis sites were aggregated at the parcel level to represent individual farms, using parcel data obtained from the National Parcelmap Data Portal (Boundary Solutions 2020). All farmed parcels were compared against a dataset of cannabis farms enrolled in the California Water Board's (CWB) statewide cannabis program, obtained via a Public Records Act request. Any parcels with cannabis cultivation that were not enrolled in the CWB state regulatory program were considered unpermitted in our analysis (n = 5,997; 406.23 ha of total cultivation area). Those enrolled in the CWB state regulatory program were considered permitted (n = 1,787; 243.06 ha of total cultivation area). All comparisons of permitted and unpermitted water use were conducted at the HUC12 watershed scale to preserve anonymity.

Zoom In Zoom Out Reset image size

We developed a modeling framework for predicting monthly water extraction volumes for cannabis cultivation in northern California, using water-use reporting information from permitted farms and geospatial data describing farm size, location, and operational characteristics (figure 2). Annual and monthly water totals were obtained from reporting data submitted by permitted cannabis farms enrolled in the CWB cannabis program. Because the data were self-reported, only farms using water meters or known application rates (i.e., emitter rates applied to watering schedules) were included in the analysis. As described by Dillis et al 2020, water extraction was considered separately from water demand. While water demand reflects the total annual amount of water a cannabis farm needs to grow their crop, water extraction reflects when this water is actually withdrawn from the environment, including water diverted to storage and directly applied to plants. Extraction volumes are therefore considered a more ecologically relevant metric of water consumption by cannabis farms. Monthly estimates of water extraction volumes were generated by first modeling annual water demand and then applying a monthly water extraction profile model to estimate how annual demand was extracted over the year (Dillis et al 2020). We then applied the models to farm mapping data for the region to estimate the monthly extraction volumes for all farms in the sampled watersheds, including both permitted and unpermitted farms. Each step in the modeling framework is described in detail below.

Zoom In Zoom Out Reset image size

Annual water demand

We first developed an annual water demand model, relying on data from permitted farms from 2018–2020 (n = 1,161). Data from CWB annual reports included cultivation area, operation type(s), and a monthly reporting of water use, which was summed for each farm to serve as the annual water demand response variable. We reviewed 2018 and 2020 NAIP aerial imagery to digitize cultivation area for each CWB farm instead of relying on reported values. Because many farms have both mixed-light and outdoor cultivation areas, we quantified operation type as the proportion of outdoor cultivation present at a site. Two additional variables were generated for each farm as predictors of annual water demand: reference evapotranspiration (ET) and terrain aspect, which were obtained from the TerraClimate dataset (Abatzoglou et al 2018), available in Google Earth Engine (Gorelick et al 2017), and Digital Elevation Models (DEMs) from the National Elevation Dataset (USGS 2020), respectively. Reference ET was calculated as an average daily value during the growing season (April - October) and was chosen to represent climatic conditions, as it encompasses surface temperature, relative humidity, and other factors that otherwise introduce collinearity if considered as separate predictors. Terrain aspect was calculated as the average parcel value (0°–360°), which was recentered to South aspect (180°) by subtracting 180 and taking the absolute value (thus South becomes 0°, East and West both become 90°, and North becomes 180°). The source data used to calculate terrain aspect were finer resolution (30 m raster cell) than reference ET (4638.3 m raster cell), thus making it more sensitive to microclimatic variations.

Four separate models were fit and compared using R statistical software (R Core Team 2018) to select a top-candidate annual water demand model (appendix A). Candidate models included: (1) a rate-based estimate (i.e., m³ m⁻²) relating cultivation area to annual water demand, (2) a simple Ordinary Least Squares (OLS) model using cultivation area and operation type as predictors of water demand, (3) an enhanced OLS model that also included reference ET and terrain aspect, and (4) a generalized linear model similar to that used by Dillis et al 2020 with cultivation area, operation type, reference ET, and terrain aspect as predictors. For each model, a training set (random subsample of 60% of the dataset) was used to make predictions for the reserved test set (remaining 40% of the dataset) and performance was assessed over 1,000 validation iterations. Although there were multiple years of data used in the analysis, training and test sets were split by farm ID so that no farms were used to make predictions for themselves in subsequent years. Model predictions for the test sets and their observed annual water demand values were compared, using percent bias, Mean Absolute Error (MAE), and r² to select the top performing model (with priority to the three metrics following the order listed). Comparisons were based on 100 iterations of training and test sets for each model.

The top performing model was the enhanced OLS regression using all of the predictor variables described above (appendix A). The model structure can be described as:

$$\begin{eqnarray}&&{W}_{i}=\alpha +{\beta }_{c}{c}_{i}+{\beta }_{t}{t}_{i}+{\beta }_{v}{v}_{i}+{\beta }_{r}{r}_{i}+{\beta }_{ct}{c}_{i}{t}_{i}+{\beta }_{cv}{c}_{i}{v}_{i}+{\beta }_{cr}{c}_{i}{r}_{i}+\varepsilon \end{eqnarray} \tag{ 1 }$$

where, W_i is estimated water demand volume of a farm (i) in cubic meters, as a function of cultivation area (c), operation type (t), reference ET (v), and terrain aspect (r), as well as interaction terms for cultivation area and operation type (ct), cultivation area and reference ET (cv), and cultivation area and terrain aspect (cr). Water demand is modeled as a linear combination of these variables, fit with slope coefficients (β) and an intercept term (α). Huber-White robust standard errors were calculated to account for heteroskedasticity (figure S1), using the sandwich and lmtest packages in R (Zeileis and Hothorn 2002, Zeileis 2006).

Monthly extraction estimation

In the second step of the modeling framework (figure 2), we partitioned annual water demand into monthly estimates based on an extraction profile model. Data from 2017–2018 water reports for permitted farms (n = 1,130) were used for determining monthly proportions of water extraction. This period predates seasonal water use regulations adopted in 2019, which prohibit the diversion of surface water for the duration of the growing season (April-October). We therefore assume that water use practices by permitted farms prior to 2019 are representative of unpermitted farms with otherwise similar operation types and site characteristics. We used the water storage capacity of farms to model monthly extraction profiles, given that storage capacity has been shown to be a primary determinant of seasonal water extraction timing (Dillis et al 2020). We normalized storage capacity by cultivation area to calculate a storage capacity to cultivation area ratio (hereafter, STCA ratio), using data from the CWB annual reports. We then fit a multinomial model to estimate the likelihood each observation of water extraction (aggregated into discrete 3.79 m³ (1,000 gallon) units for each monthly reported total) occurring in a given month (M), based on the STCA ratio (s) of the farm (i) at which it occurred. The multinomial logit structure of the model ensured that all monthly likelihood estimates summed to unity. Monthly likelihoods were therefore interpreted as the monthly proportions of total annual water extraction. A generalized linear model was fit using the nnet package (Venables and Ripley 2002) in R statistical software. The model structure can be described as:

$$\begin{eqnarray}&&mlogit\left({M}_{i}\right)=\alpha +{\beta }_{s}{s}_{i}+\varepsilon \end{eqnarray} \tag{ 2 }$$

where an intercept term (α) is added to the storage ratio term (β_s ) to produce a log-odds estimate of the likelihood of a water extraction observation occurring in a given month. Log-odds estimates were converted to monthly likelihoods prior to multiplying them by the annual demand estimates in order to generate monthly extraction volume estimates. The predictive performance of the entire composite model was analyzed on a monthly basis, using the same process and criteria (MAE, percent bias, r²) as described for the annual demand models.

We next applied the water demand models to all farmed parcels in the study watersheds (figure 1). For each farmed parcel (permitted and unpermitted), we calculated the cultivation area and operation type variables using 2018 NAIP aerial imagery. We also calculated reference ET and terrain aspect for all farmed parcels, as described above. We then used the trained model from equation (1) to predict annual water demand at all farms with the calculated explanatory variables. To estimate the monthly extraction volumes of cannabis farms, we first determined whether an irrigation pond was present and estimated the likelihood that there was access to groundwater at the site. These factors have been previously shown to strongly influence the seasonal patterns of water extraction at cannabis farms (Dillis et al 2020). Specifically, farmers who rely on groundwater wells tend to employ limited or no long-term water storage and thus have water extraction profiles that track the water needs of cannabis (i.e., water is extracted and applied on demand). In contrast, farmers who rely on surface water tend to have more water storage, as these sources tend to be more ephemeral (and are subject to seasonal use restrictions), and therefore extract water in the winter months, outside of the growing season. Farms that have access to large on-site storage, such as an irrigation pond, extract most of their water in the wet season, such that water extraction is out-of-phase with plant demands (Dillis et al 2020).

Aerial imagery for each farmed parcel was inspected to determine the presence of an irrigation pond. To determine whether or not farms had access to groundwater, we used a classification model described by Dillis et al (2021b) that relies on water source information from permitted farms (n = 1,237) in Humboldt and Mendocino County. The model used 13 variables describing the local topographic, climate, and watershed characteristics of farm locations (table S3) to make a categorical prediction of farm reliance on groundwater (i.e., well/no well). All farm-level model parameters were generated from open-source GIS data, including Digital Elevation Models (DEMs) from the National Elevation Dataset (USGS 2020), watershed boundary and stream flowline data from the National Hydrography Dataset (USGS 2019), and precipitation data (30-year annual average) from the PRISM Climate Group (PRISM Climate Group 2018). The model was trained with source-water information at permitted sites and tested against CDFW inspection data from unpermitted farms for which a water source was recorded (n = 80), which demonstrated an acceptable level of accuracy (71.75%) and bias (+2.65%).

Based on the observed presence of irrigation ponds and predicted well use, all farms in the study area, including permitted and unpermitted farms, were assigned to one of four categories: (1) those with wells, but without ponds, (2) those without wells and ponds, (3) those without wells and with ponds, and (4) those with wells and ponds. Each of these farms was then assigned a STCA ratio based on the mean values calculated for permitted and unpermitted sites within each respective category, using CWB reporting data and CDFW inspection data. STCA ratio values were then used in the water extraction profile model to estimate the monthly proportion of water extraction, and subsequently applied to estimated annual demands, to produce monthly extraction volumes at all farms in the study area. Monthly water extraction volumes for permitted and unpermitted farms were then aggregated by watershed (HUC12).

To provide context for cannabis water extraction estimates, monthly instream flow estimates were calculated for each study watershed using a natural flow model described by Zimmerman et al (2018). Mean monthly natural flow estimates were generated for each outlet stream segment of study watersheds, converted from cubic feet per second to cubic meters per month, and compared against cannabis water extraction estimates. We specifically evaluated water extraction volumes relative to natural water supply volumes during the peak growing season (July—September).

Finally, we used our modeling framework to explore how dry season (July-September) water extraction demands for the region might change under alternative future policy scenarios. We considered hypothetical scenarios to compare the efficacy of management approaches used to reduce summer water extraction: prohibiting surface water diversions in the dry season, expanding storage capacity (to allow for capture of more water in the wet season), and implementing both strategies simultaneously. The first policy scenario, Full Compliance, considers a future in which unpermitted farms become permitted and all cannabis cultivators comply with seasonal forbearance regulations under existing policy, which prohibits surface water diversions between April and October. For this scenario, all unpermitted farms that rely on surface water were assigned to a monthly extraction profile in which all water demands are met by surface water diversions in the wet season (i.e., by storing water). A second scenario, Median Storage, was used to represent a future in which all cannabis farms (permitted and unpermitted), regardless of water source, expand their storage capacity to the median STCA ratio value reported by permitted farms. This represents a plausible level of potential increased storage capacity for farms, while recognizing the logistical and financial challenges of storage expansion (Dillis et al 2020). To represent a less ambitious storage target, we also considered a 25th Percentile Storage scenario, in which all farms have access to storage equivalent to the 25th percentile of the STCA ratio of permitted farms. A final scenario, Combination, was also considered to represent a future in which all farms comply with the forbearance requirement but also expand their storage capacity to the median STRA ratio value. We re-ran the models with the revised farm attributes from each scenario and calculated the percent change in dry season water extraction volumes for all watersheds in the study region.

The OLS model with reference ET and terrain aspect performed best in predicting annual water demand relative to the other model configurations (MAE = 303.12 m³; Pbias = 0.10%; r² = .34; table S2). The enhanced OLS annual water demand model (table 1) showed reliable interactions between cultivation area and the three other predictor variables operation type (MLE = −0866; SE = 0.0292), reference ET (MLE = 0.003; SE = 0.001), and terrain aspect (MLE = −0.016; SE = 0.003), indicating that their effects are amplified with increasing values of cultivation area (figure 3). Although a simpler OLS model did not perform as well in predicting annual water use, the model coefficients are included as supplementary material (Table S4). This model provides a method for estimating annual water demand from a cannabis farm that does not require processing reference ET or terrain aspect data, yet performed only slightly worse than the enhanced OLS model and was superior to the rate-based model (table S2).

Zoom In Zoom Out Reset image size

In the extraction profile model, variation in the STCA ratio resulted in strongly divergent monthly extraction patterns (figure 4), consistent with Dillis et al 2020. Coefficient estimates (table 2) demonstrated that relative to January (reference level), the summer months of July (MLE Intercept = 2.08, SE = 0.03), August (MLE Intercept = 2.23. SE = 0.03), and September (MLE Intercept = 1.97, SE = 0.03), were reliably predicted to experience the greatest proportion of annual water extraction. Coefficient estimates for the STCA ratio were reliably negative and most strongly negative in the summer months (July MLE = −13.38, SE = 0.27; August MLE = −14.48, SE = 0.28; September MLE = −13.77, SE = 0.29), reflecting that with increasing amounts of water storage, less water is extracted during this period of the year. Of the four farm typologies, those using a well and that did not have a pond for storage had the lowest STCA ratio and the highest predicted proportion of water extraction in July (18%), August (20%), and September (16%). By contrast, farms without wells, but with an irrigation pond had the highest predicted proportion of water extraction in January (21%), February (22%), and March (21%).

Zoom In Zoom Out Reset image size

Predictive performance of the composite model (annual demand and monthly extraction) varied across months and was generally best during the growing season (May through October; table 3). While MAE was mostly consistent across most months, percent bias was notably lower in July (6.00%), August (3.20%), and September (4.80%) than in the off-season months of November (109.00%), December (96.90%), and January (79.90%). The r² values were also higher in the months of July (r² = 0.36), August (r² = 0.35), and September (r² = 0.34) than in November (r² = 0.08), December (r² = 0.11), and January (r² = 0.14). This discrepancy in model performance was likely due to the fact that water extraction in the offseason is being input to storage and can vary by farms based on local water availability and diversion schedules. Additionally, offseason water extraction (November through April) often involves input to irrigation ponds, the capacities of which can include extreme outliers that the model cannot account for. Predictive performance of the composite model in the offseason months was considerably improved as outlier values of reported water extraction were removed (appendix A).

Farm-level estimates for all mapped permitted (n = 823) and unpermitted (n = 5,959) farms, generated by the composite water extraction model, were aggregated at the watershed scale. The number of unpermitted farms (Median = 9; IQR = [3, 47] and associated cultivation area within watersheds (Median = 0.021 km²; IQR = [0.006, 0.047) was typically greater than permitted farms (Median number = 1; IQR = [0, 7]; Median size = 0.003 km²; IQR = [0.000, 0.013]). There was also noticeable spatial divergence between watersheds in terms of permitted and unpermitted cultivation (figures 5(A) and (B)), with some primarily supporting permitted farms and others primarily supporting unpermitted farms. The distribution of overall (permitted and unpermitted) water extraction totals reflected these differences and varied by orders of magnitude between watersheds (figure 5(C)) and primarily tracks the pattern of unpermitted cultivation given its larger presence on the landscape. That is, the watersheds with the highest annual water extraction totals are those with the largest footprint of unpermitted cannabis cultivation.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Differences in extraction profiles of permitted and unpermitted farms were most pronounced for farms without wells or ponds. Unpermitted farms of this typology had STCA ratio values (0.01) much smaller than permitted farms (0.21), demonstrating more on-demand extraction during the summer months (figure 4(A)) by the former. By contrast, there was little difference between permitted and unpermitted farms that relied on groundwater, as both had relatively low STCA ratios (0.03; 0.01), and thus extracted water on demand. As a result, differences between permitted and unpermitted water extraction patterns were amplified in areas where farms were predicted to rely more on surface water than groundwater. Differences between permitted and unpermitted extraction volumes were most pronounced in the dry season (figure 7). For example, differences in July (Median = 1120.83 m³ km⁻²; IQR = [356.05, 2898.83]), August (Median = 1285.48 m³ km⁻²; IQR = [411.40, 3327.11]), and September (Median = 1002.77 m³ km⁻²; IQR = [319.43, 2594.22]) were much larger than in November (Median = 166.67 m³ km⁻²; IQR = [45.71, 453.66]), December (Median = 261.69 m³ km⁻²; IQR = [63.32, 691.66]), and January (Median = 351.54 m³ km⁻²; IQR = [90.39, 954.05]).

Zoom In Zoom Out Reset image size

Most of the water extraction associated with unpermitted farms was concentrated in a small subset of watersheds (figure 8). Almost one third of all unpermitted water extraction was attributed to 10 of 115 study watersheds (8.6%) on both an annual basis (32.93% of total unpermitted extraction) and during the peak water extraction period (July - September; 33.21% of total unpermitted extraction). The unpermitted water extraction in these 10 watersheds also accounted for over one quarter of all (permitted and unpermitted) estimated water extraction in the study area on both an annual basis (26.59%) and in the peak water extraction period (28.53%).

Zoom In Zoom Out Reset image size

We found that cannabis water demand for the peak growing season (July—September) was estimated to represent a small percentage of naturally available surface-water supplies, when evaluated at the HUC12 watershed scale (figure 6). Estimated percentages did not exceed 1% of estimated instream flow in 89 of 115 watersheds and only exceeded 2% in 12 of 115 watersheds. The largest estimated value for a single watershed was 13.53%. Spatial patterns in cannabis water extraction expressed relative to water availability tracked the distribution of unpermitted cannabis (figure 5(D)), reflecting the disproportionate amount of water used by unpermitted cannabis in the dry season.

For the Full Compliance scenario, in which all farms in the study area were assumed to comply with seasonal surface water diversion restrictions, water extraction during peak summer months was reduced in most watersheds (July-September; figure 9). Water extraction was reduced by 50% or more in 34 watersheds under the Full Compliance scenario. However, in the majority of watersheds (63 of 115), dry season water extraction was reduced by less than 25% relative to the current extraction estimates. This is because many of the farms in these watersheds rely on groundwater wells and are not subject to dry season forbearance requirements. Enhancing water storage capacity had a more significant effect on dry season extraction estimates. Under the Median Storage scenario, 112 of the 115 study watersheds were estimated to have water extraction reductions of 25% or more, with estimated reductions of 50% or greater in 96 of 115 watersheds. Notable reductions in extraction volumes were also estimated under the 25th Percentile Storage scenario, in which 103 of 115 watersheds were estimated to have reductions of 25% or more, with estimated reductions of 50% or greater in 60 of 115 watersheds. The Combination Scenario, which assumed farms would both enhance storage and forbear from surface water diversions in the dry season, produced the largest reductions in summer water extraction, with 100 of 115 watersheds having reductions of 50% or more and 112 of 115 watersheds having reductions of 25% or more.

Zoom In Zoom Out Reset image size

Unpermitted water extraction is responsible for a disproportionately large share of total water consumption by cannabis agriculture in Northern California and this difference is magnified in the driest months of the year when streamflow is lowest (Bauer et al 2015, Carah et al 2015). Based on estimates of water extraction relative to instream flow, our results indicate that cannabis water demands are likely more of an issue at the scale of smaller headwater streams than over an entire watershed. Permitted farms typically have access to water storage to collect irrigation water in wet season months, thereby (at least partially) limiting surface water diversions in drier periods. However, our results suggest that storage capacity of unpermitted farms is more limited, resulting in relatively higher dry season extraction proportions. Given the higher number of unpermitted farms in the region and greater cultivation footprint, dry season demands by unpermitted farms in the region remains a potentially significant pressure on water resources in the region.

Our scenario analyses suggest that there are opportunities to reduce the amount of water extraction from watersheds in the dry season, including transitioning unpermitted farms to the regulated industry to comply with seasonal forbearance requirements as well as the expansion of on-site water storage capacity at both permitted and unpermitted sites. The spatially clustered nature of unpermitted farms suggests that highly impacted watersheds may be valuable targets for outreach efforts aimed at transitioning unpermitted farms to the regulated industry. Campaigns to increase water storage capacity are also likely to be helpful in reducing dry season water demands at both permitted and unpermitted farms. Our scenario modeling demonstrated that the most effective and widespread reduction in summer water extraction would occur if all farms implemented a moderate amount of storage. An additional avenue by which water extraction may be reduced, but not examined in this study, is through improvements in water use efficiency, thereby decreasing demand. The broad range of annual water use values (normalized by cultivation area) we observed suggest that some cannabis farms are using substantially more water than others. Horticultural research to produce a crop coefficient for cannabis, as exists for nearly every other agricultural crop, would give farmers a better sense of at what point crops are being overwatered and no longer using all of the irrigation water they are allocated.

A robust assessment of cannabis water-use impacts on streamflow requires improvements in estimation of instream water availability, characterization of subsurface hydrogeologic properties, and quantification of other anthropogenic water uses. The impacts of cannabis cultivation are most likely to occur in small headwater streams (Bauer et al 2015), which rarely have gaging stations, making it difficult to directly measure streamflow responses. However, recent advances in the use of machine learning models (e.g., Grantham et al 2022) and other hydrologic modeling methods have improved the accuracy in predicting flows in ungaged streams. The strategic placement of new gages, coupled with improved models, would be helpful for developing reliable estimates of streamflow in cannabis-producing watersheds. This study reinforced previous findings by Dillis et al 2021 documenting that many cannabis farmers in Northern California are relying on groundwater wells to meet their irrigation needs. Extraction from groundwater wells can have similar effects as direct surface water diversions, although the timing and magnitude of streamflow depletion is strongly affected by subsurface soil properties (Zipper et al 2019). For example, Hahm et al 2019 demonstrated that watershed lithology can have a significant effect on subsurface water storage and streamflow characteristics in Northern California streams. In particular, watersheds underlain by weathered bedrock (agrillite and sandstone) in the Coastal Belt have greater storage capacity than those dominated by the Central Belt mélange (argillite matrix) (Langenheim et al 2013, Hahm et al 2019). As a result, streams overlaying Central Belt lithology tend to be flashier, tracking patterns of seasonal rainfall, and often dry in summer, which could make them more sensitive to streamflow depletion from water diversions. In contrast, Coastal Belt streams are often sustained by subsurface flows throughout the dry season (Hahm et al 2019) and the greater storage capacity could make them less sensitive to both groundwater and surface water withdrawals.

In addition to considering the irrigation water source and hydrologic properties of the watershed, estimation of streamflow impacts in cannabis-producing watersheds must consider the influence of other water users, which have demands that often exceed those of cannabis. For example, Zipper et al 2019 reported that water extraction for residential use contributed more to streamflow depletion than cannabis in another northern California watershed. Finally, to improve estimation of potential streamflow impacts additional data from both permitted and unpermitted farms are also needed to better characterize water use efficiency, storage capacity, and water extraction volumes. For example, field observations of unpermitted farms by CDFW staff frequently document sources of water waste including leaking pipes, overflowing passive inputs to water tanks, and unlined irrigation ponds. Understanding the extent to which these types of waste characterize water use practices is critical for developing water-use efficiency standards and promoting strategies to limit streamflow depletion risk.