We therefore examine this cycle not only as a means to save water, but ask if and how it can enhance the viability of nonindustrial farming operations as the food system adapts to restricted water availability. We consider the relevant policy recommendations outlined in Blesh et al.’s analysis of how institutional pathways can act synergistically with farmer networks to enable agricultural diversification , asking which have the potential to point future dry farming towards scaling size vs scope.To better situate these policy options in the local context, we first look to the outcomes of institutional intervention in organic strawberry production in a very similar region on the Central Coast, and consider the analogous options for dry farm tomatoes. Similar to dry farm tomatoes, organic strawberry production was launched into the spotlight by government-mandated input curtailments . For strawberries, the development of an organic strawberry production system also coincided with the adoption of an organic certification process by the US Department of Agriculture. Growing public interest in organic strawberries and the methyl bromide ban led to the rapid expansion of industrial-scale organic strawberry production– blatantly scaling size of production . As production increased, organic strawberry markets saturated and prices crashed, leaving an economic landscape where only the largest operations could remain viable selling strawberries at market prices . At this point, agroecological growers had to redouble their efforts to target local consumers with direct marketing strategies, weed drying rack as the organic label no longer added the necessary value to profitably sell their product.
In an analogous case for dry farm tomatoes, it is easy to see the immediate appeal of establishing a “dry farm” label that can incorporate the social value added to dry farm tomatoes into the price of the product without relying on consumers trusting and paying a premium based solely on higher qualities. However, by divorcing dry farm practices from quality premiums and trusting relationships with customers, a dry farm label would make it much easier for large-scale growers to enter the dry farm market. These larger operations–which may struggle to produce high quality fruits or maintain direct relationships with customers but can still decrease water usage enough to produce a certified dry farm tomato–could easily grow dry farm produce at large enough scales to edge smaller growers out of the label. As has been seen in the organic program, industrial growers could also lobby for an official relaxation–a literal watering down–of label standards . This sidestep of the dry farm practices described in the above interviews would not only further advantage large scale farmers, but would also undermine the very water savings that they are meant to encourage.Larger scale growers may also be favored when farmers are paid to implement specific practices. Administrative costs involved in enrolling in payment-for-practice programs can be a cumbersome barrier to entry, while low payouts at small scales dissuade small farmers who implement the practice from enrolling . These patterns are currently seen in programs offering cost shares for cover cropping, where farm size is significantly larger for participants than non-participants .Given farmers’ interest and current experimentation with dry farming non-tomato vegetables, expanding the set of crops that can be dry farmed and adapted to local conditions is a clear target for future policies. Support for research and participatory breeding programs/variety evaluation could spur development of locally-adapted dry farm varietals.
By compensating farmers for experimentation with diversified dry farm rotations and development of locally adapted varietals, policymakers can also absorb some of the risk inherent to on-farm experimentation and encourage innovation on the farms that are most familiar with the practice, while simultaneously lowering barriers for farmers new to the practice. To create a policy environment where experimentation feels more accessible to farmers, minimum lease terms could be set for farmland, allowing farmers to feel more secure in investing in localized practices .Priority could also be given to creating programs that connect farmers–particularly new farmers and those who hold underrepresented identities–to available farmland. Without the burden of securing water access, lands that would otherwise be impossible to farm with summer crops could become arable, particularly in conjunction with the concurrent support of the other policies discussed here. Though many areas will still require some access to water to successfully dry farm , crops’ need for water coincides with points in the season when surface water is most available , making areas with inconsistent water access over the course of the season likely candidates for dry farm success. Priority might initially be given to areas shown as suitable on the map, but as new and locally adapted crop varieties emerge, access may also extend.What is the difference between data and information? Although these two terms are often used interchangeably in biology, entire sub-fields of data science are dedicated to solving this riddle . While mathematicians have developed formal definitions for these terms , we will attempt to illustrate the conceptual differences with a simple example. Suppose a graduate student wants to analyze the gait dynamics of a dairy cow using a 120 Hz leg-mounted accelerometer.
In developing their data collection protocol, they surmise that if one accelerometer would be good, then two accelerometers must be even better. For their pilot study, they strap two sensors onto the same leg of their favorite cow and record her walking normally down the alley way in a straight line. If both sensors were calibrated and synchronized properly, then these sensors might produce data that looks something like the plot in Figure 1A. From this graph, can we learn anything about this cow from the blue sensor that we could not have inferred from the red sensors ? Since these sensors are quite precise, the two signals are virtually identical. Thus, even though this graduate student has doubled the number of data points recorded by adding a second sensor, they have not really collected any additional information about the gait dynamics of this cow.As this example has hopefully highlighted, data can be considered a physical resource ± something that can be measured in bytes and memory space. Information, on the other hand, should be regarded as a more nebulous unit of measure that represents how much can be learned from a dataset. In experimental settings, collecting measurements is typically costly, so sampling strategies are often developed to minimize redundancy between data points, such that data and information are often functionally equivalent terms . With sensor technologies, however, sampling density is often dictated by the hardware, which can result in a considerable amount of redundancy between data points . With such datasets, information compression strategies are often implemented in an effort to reduce the size of a dataset without losing any potentially useful information about the system. Returning to our previous example, suppose the graduate realizes they have not requisitioned enough hard drives to store data from both accelerometers. They decide to average the output of the two sensors at each time point, as illustrated in Figure 1B. This simple information compression step reduces the size of their dataset by half, but no details about the signal have been sacrificed. In the struggle that ensues to attach the accelerometers, one sensor is accidentally strapped on upside down. As a result, the two signals become inverted, as illustrated in Figure 1C. Subsequently, if the previously validated rolling average filter were applied to this new dataset, mobile rack the signals would cancel out, leaving only the measurement noise as shown in Figure 1D. Thus, this example serves to illustrate just how easily biologically relevant patterns can be lost when the assumptions employed in an information compression strategy do not match the data to which it is applied. As the range of Precision Livestock Farming technologies that are available to farmers to herds continues to expand , sodoes the need for robust algorithms that are capable distilling the large quantities of data that these systems produce down to useable knowledge . In data science, there LV ³QR fUee OXQcK´ ±no single algorithm that can be expected to perform optimally in all scenarios . In this paper, however, we will demonstrate how model-free unsupervised machine learning approaches can be used to recover and encode complex behavioral patterns from datasets generated outside of controlled experimental contexts, where an appropriate model may not be immediately obvious.
We will then explore how model-free information theoretic approaches may be used to recover complex nonlinear associations between farm records and multi-sensor systems in order to provide more comprehensive ethological insights.Animal scientists are seldom forced to work with raw accelerometer data, as in the previous example. Therefore, to further illustrate the fundamental differences between model-dependent and model-free approaches to information compression and knowledge discovery, lets consider a more practical example. Suppose a farmer has a group of 100 cows that are currently overstocked at 200% stocking density in a free stall barn. Concerned they might fail the outcome measures on their upcoming welfare audit, the farmer contracts a consultant to analyze data from the leg mounted accelerometer system that the farm uses in its estrus detection program.The farmer would like to know if there is any clear evidence of animals with compromised welfare in 60 days of archival records? They begin by implementing standard exploratory data analysis techniques, and produce the plot in Figure 2, wherein each dot represents the proportion of time a given cow is recorded lying down on a given observation day, with a LOESS curve fitted to reflect the mean lying time for the entire herd at each time point. From this visualization they glean that there is a considerable amount of variation between cows, and that there certainly are animals on any given day that are not spending a sufficient amount of time lying down . There is not, however,any clear evidence in this graph that would indicate that lying time patterns are changing at the herd level over time – quite the opposite actually, the herd mean is incredibly stable over the observation interval.Based on this visualization, the consultant constructs a mixed effect model with ³cow´ as the random effect, to account for repeated measurements and avoid pseudo-replication. Since the data appears stationary, they add ³day of observation´ as a simple continuous fixed effect. Finally, based on their preexisting knowledge of factors that may affect lying time, they add a categorical fixed effect to distinguish between ³heifers´ and ³cows´ within the herd. The results of this modelare provided in full in Supplemental Materials. As anticipated from the EDA visualizations, the fixed effect for day is neither biologically nor statistically significant . The consultant is, however, a bit surprised to find that the difference in average lying time between heifers and cows is also neither statistically or biologically significant . Thus, the consultant determines that the expected proportion of time spent lying for any animal within this herd for any given day within the observation window is 40% . In appraising the variance estimates for the random effects in this model, the consultant also sees that between-cow variance is considerably smaller than the within-cow error term , from which they conclude that there is also no evidence of consistent individual differences in lying time within this herd, and so these results also do not indicate that any individuals are consistently above or below the expected lying time for this herd. Subsequently, they might confirm that, while cows might not get to lie down an adequate amount of time every day, and that the average lying time could certainly be improved, there is also no evidence of cows in this herd that are consistently, and critically, under-rested. Let¶ s now suppose that the milk buyer is dubious of this result, and so they send the data to a new consultant who specializes in welfare audits. They ask the farmer some additional questions about this herd before constructing their own model and learn that the start of grazing season fell somewhere in the middle of this observation window, at which point cows had free access to pasture from their free stall barn. This consultant, suspecting that heifers and cows might react differently to this pasture access, develops a similar model to the first consultant, except this time an interaction term is included to allow the temporal dynamics to differ between these subgroups. With this small modification to the fixed effects matrix, this model now tells an entirely different story.