The risk of contamination, thus, typically includes the cost of withdrawing product from the market and lost revenue from a portion of harvest. Larger operations, therefore, have greater risk from contamination, though the risk from forgone revenues and the associated costs of product recalls are presumed to increase proportional to sales. This component of risk is likely to be scale-neutral, suggesting that the optimal private investment in food safety is independent of whether a product is produced in a concentrated industry or by many small firms. But large firms are more likely than small operations to have brand capital, which is threatened by food scares traced to their products. The loss of a firm’s good reputation can occur as the consequence of a food scare independent of the magnitude of the scare and the firm’s market share. With the loss of reputation, a brand-name product would lose its price premium. Sales and margins for products produced by the firm unrelated to the food scare are also likely to suffer.Large firms with good reputations, therefore, stand to incur losses from lapses in food safety that are disproportionately higher than small firms. Losses to a firm from an outbreak of food borne illness often also include product liability for related illness and death. Judgments can easily reach into tens of millions of dollars. If food contamination occurs in the field, then the magnitude of the outbreak may be independent of the market share of the responsible producer. However, if contamination were to occur in a processing facility,hydroponic vertical farming then the greater quantities handled in the facility and distributed through a wider network could cause illnesses and fatalities to be greater for outbreaks caused by larger firms.Liability is limited to the assets of the firm; thus, it increases as production increases because the assets of larger producers are greater.
Limited liability explains, in part, why firms that operate in hazardous industries with latent damages tend to be smaller than firms in other industries. Divestiture is recognized as a mechanism to limit liability. Small growers, then, face an upper bound on the risk of food contamination that lowers their incentives to invest in food safety. For instance, the Colorado cantaloupe grower whose melons are implicated in 32 deaths filed for bankruptcy in May 2012, listing its net worth at -$400,000. The farm’s owners, therefore, avoid potentially tens of millions of dollars in liability. Because large firms generally have more assets, they face greater exposure from product liability, and, therefore, demand greater protection against food contamination. Large firms and small firms alike can insure against product liability, but premiums are positively correlated with coverage limits and with past lapses in food safety, so that large firms demand greater ex ante prevention of food contamination even if they are insured. Large firms face more-than-proportionally greater risk from selling tainted food, and, consequently, have greater demand for food safety. They also face lower average variable costs in supplying food safety. Therefore, the optimal level of investment, which equates the marginal benefit of an incremental increase in food safety with the marginal cost of achieving it, is increasing at an increasing rate in firm size. Holding market supply of food constant, then, the provision of food safety declines as the number of producers grows. Empirical evidence from the meat and poultry-processing sectors indicates that firm size is an important determinant of firm-level investment in food safety. For instance, research suggests that firm size has more impact on the adoption of safety and quality assurance practices than any other firm characteristic. Large firms are also more likely to have adopted a range of food safety technologies.The losses stemming from an outbreak of food borne illness often are not limited to those firms implicated in food contamination. Indeed, food safety regulators issue broad warnings about food products irrespective of where they originated if the origin of contamination cannot be immediately identified.
For instance, when in 2006 an outbreak of E. coli was linked to consumption of bagged fresh spinach, the FDA issued a blanket warning to consumers to avoid the product altogether. Fifteen days later, the alert was scaled back to include a warning against consumption only of specific brands of spinach packaged in California on specific days, but the industry had already experienced a dramatic decline in sales. The outbreak, blamed for 199illnesses and three deaths, cost spinach producers $202 million in sales over a 68-week period, a 20% decline. By 2008, demand for California spinach remained below pre-outbreak levels. Because a food scare can induce a negative demand shock, there exists a reputational externality associated with food safety provision, which, consequently, exhibits public good characteristics. The benefits of an individual firm’s food safety investments accrue, in part, to competing firms. Because the firm does not capture the full benefits of its food safety provision, it will under invest in food safety relative to the efficient level. Industry-wide private provision may be much too low. A firm’s share of the benefit from an investment in food safety, however, is increasing in its market share. The more concentrated the industry, therefore, the closer is the equilibrium level of food safety provision to the efficient level. A monopolist would internalize the full benefit of his investments, and, therefore invest in the efficient level of food safety. As the number of small firms increases, however, the equilibrium market-wide provision of food safety falls increasingly short of the efficient level.Because agronomic and climate conditions impact the optimal handling and processing of crops, they afford some firms and regions a comparative advantage in producing safe food. For instance, the hot and dry conditions during the cantaloupe-growing season in California reduce the crop’s exposure to contaminants that can be transferred to melons in wet fields. Moreover, because the dry conditions keep California cantaloupe relatively clean, most are packed directly in the field, requiring less handling and avoiding exposure to food pathogens in shed packing operations that rinse and dry the produce.
As retailers seek to market local produce, however,vertical hydroponic garden comparative advantage in the provision of food safety is forsaken. The Listeria outbreak last summer was linked to unsanitary conditions at the Jensen Farms packing shed. And an FDA investigator identified unsanitary practices at the Chamberlain Farms packing shed that has been associated with this summer’s Salmonella outbreak. Concentration of production in regions with comparative advantage creates agglomeration advantages for the mutual provision and certification of food-safety practices. Because of the potential losses from food scares and the market-wide externalities from food safety investments, grower organizations have adopted voluntary process standards to mitigate risk and avoid shirking among their members. Some growers have also created marketing orders to enforce handling practices and require audits of all operations covered by the agreement. The California Leafy Green Product Handler Marketing Agreement was implemented in 2007, following the 2006 E. coli outbreak linked to spinach from California’s Salinas Valley. The California Cantaloupe Advisory Board responded to last summer’s Listeria outbreak by imposing mandatory certification by state auditors of all growers in the state. Such industry-wide cooperation and self policing is likely to be lost when production is fragmented and spread across wide geographical areas in the quest for local production.Irrigated agriculture exerts strong controls on global food production yields and the water cycle while accounting for 85-90 percent of human consumption . However, little is known about the spatial distribution of agricultural fields, their crop types, or their methods of irrigation. Spatially explicit knowledge of these field attributes is necessary in order to implement more water-efficient agricultural practices and plan for more sustainable economies . However, most remote sensing based mapping efforts cover limited political boundaries, on the order of U.S. states or smaller, and usually only cover a snapshot in time. Agriculture maps in developing countries are even more lacking in semantic detail, coverage, and resolution, which is a particularly acute problem given that agricultural expansion in these regions tends to be decentralized and without a guiding management plan for water sustainability. While a fine scale and up-to-date census of global agriculture does not yet exist, it is feasible that we can map a subset of fields that are spectrally and visually distinct from surrounding land cover and numerous enough to train a machine learning model to map a substantial subset of global agriculture. Center pivot irrigation fits these criteria; they are relatively uniform in shape, have a narrow range of sizes within particular geographies, and in drylands, can be strongly contrasted with the surrounding lack of vegetation .
Like all remote sensing applications for agriculture detection, image obstruction by clouds, fallow periods and the growth of non-agricultural natural vegetation in the landscape all pose challenges for models that detect center pivots. In more humid environments, center pivots have distinctive growing season patterns and larger amplitude changes in vegetation “greenness” compared to other vegetation types, making them readily identifiable with time series of multi-spectral images. However, it’s difficult to assemble a comprehensive time series of imagery over many parts of the world, primarily due to cloud occlusion and scene availability. Scene availability is particularly low for dates prior to the launch of the Sentinel-2, so a method to map center pivots using single date imagery in many parts of the world is desirable. Such a method is particularly valuable given that center pivots are one of the most ubiquitous irrigation sprinkler systems employed in large scale commercial agriculture and make up a considerable fraction of the unplanned agricultural expansion in developing countries. There are a variety of methods for mapping objects using remotely sensed imagery, but object based classification has been shown to outperform per-pixel classifiers in cases where the object of interest contains enough pixels to distinguish objects by textural or shape properties . Because center pivots are large relative to the resolution of public sensors like Landsat, they are amenable to being mapped using object based classifiers. The traditional approach to object based classification in remote sensing has been to use manually tuned algorithms to delineate edges or engineered features that capture texture and shape properties combined with per-pixel machine learning algorithms . However, traditional object based classifiers in remote sensing have a tendency to over fit and must be manually tuned or supplemented with region specific post processing to arrive at a suitable result. Another class of methods which make use of convolutional neural networks have achieved great success on complex image recognition problems in true color photography. These have not been thoroughly evaluated for mapping agricultural fields as instances across a large climate gradient using Landsat imagery. The goals of this research are twofold. First, I evaluate the performance of convolutional neural network based instance segmentation models on Landsat imagery. This experiment determines if CNN based models can make use of Landsat’s 30 meter resolution to provide reliable predictions of the locations and extents of center pivot agriculture in various states of development, including cleared, growing, and fallow. I test this approach by using the current most popular and near state of the art Mask R-CNN model, an approach based on a lineage of regional CNNs which jointly minimize prediction loss on region proposals, object class, refined object bounding boxes, and an object’s instance mask. This model is tested on the 2005 CALMIT Nebraska Center Pivot Dataset, which is divided into geographically independent samples that were partitioned into a training, validation and testing set. Multiple model runs with varying hyperparameters and preprocessing steps were conducted to arrive at the most accurate result on the validation set, and the final most accurate model was applied to the test set to produce the final reported accuracy. The model is also evaluated on the full training data set and 50% of this dataset in order to examine the effect of reduced training data on model accuracy over a large region. Second, I compare these results to the Fully Convolutional Instance Aware Segmentation model, which was previously the state of the art in instance segmentation in true color photography prior to Mask R-CNN.