My goal is to determine whether Mask R-CNN and FCIS produce substantially different results when applied to the task of segmenting center pivots, and if these differences are as pronounced as their performance on segmentation tasks in true color photography. This can inform the field of remote sensing on the value of adopting new CNN based architectures and to what extent performance on datasets like COCO translates to the remote sensing domain. The FCIS model is tested on the same Nebraska dataset with minimally altered hyperparameters based on Rieke , which applied FCIS to the task of segmenting large agricultural fields in Denmark. It is evaluated on a validation set, as the validation set was not used to make decisions to change hyperparameters. Both the FCIS model and the Mask R-CNN model performances on the Nebraska dataset are compared relative to their results on Common Objects in Context, a large image dataset similar to Imagenet but with added annotations for segments of objects. Imagenet is a dataset with 3.2 million images labeled with categories appearing in the images, while COCO or Common Objects in Context, contains 328,000 images with 2.5 million labeled instance segments . Answering these questions can help inform users of these models of their compared to one another and whether differences in performance are consistent across image domains. The ultimate goal of this research is to improve the mapping of field boundaries that are either actively cultivated, recently cultivated,vertical home farming or soon to be cultivated using the public Landsat record.
The dataset used does not allow for classifying particular crop types, though the Mask R-CNN method allows for this. Nevertheless, accurate field boundaries obtained from CNNs could make it much easier to classify crop type by allowing crop type classification procedures to focus on known field locations. The popularity of center pivots stems from their low energy consumption relative to the amount of water applied, low maintenance costs, and uniformity. These advantages make center pivot irrigation productive in both temperate regions and drylands with access to groundwater. However, while center pivots have been highly productive, their unchecked expansion in water-scarce regions has contributed to water shortages, over-tapped aquifers, and even political conflict. These risks are exemplified by Saudi Arabia in the 1970s and 80s, when center pivot development expanded and into the present day, where domestic production has slowed and must eventually decline . Like many of the world’s most agriculturally productive regions that depend on fossil water from aquifers, Saudi Arabia’s overall rate of irrigation has long since eclipsed groundwater recharge rates back to the Arabia Aquifer. As early as 1951, it was noted that the first government agricultural project at Al Kharj was dependent on ancient aquifer waters, and that parts of the project were already “… operating close to the margin of safety in regard to the current water supply.” . Since then, the government of Saudia Arabia has transformed large tracts of it’s deserts into fields of center pivots , resulting in a tripling of total water use between 1980 and 2006, 936% of total renewable water . This boom in irrigated area has been followed by a curtailing of domestic production, with the government pledging to eliminate water intensive wheat production by 2016 , though this commitment has since been postponed. Though researchers have called for more investment in demand side controls on water resource sustainability, the government has invested heavily in direct foreign investments in water intensive agriculture abroad.
Some of these investments are ongoing in areas experiencing extreme water scarcity, including southern Arizona, which depends on declining flows from the Colorado river, and Gambela, Ethiopia, where a government mandate to relocate smallholder farmers for a Saudi agriculture project led to armed conflict . There is a lack of public information on the distribution and ownership of large scale commercial agriculture, particularly in developing countries. Rectifying this knowledge gap will allow us to make more informed decisions of how to allocate water resources to best serve ecosystems, industry, and local populations. In regions where this information is available, it can enable improved monitoring of water rights compliance and experimentation with potentially more sustainable water management practices. A potentially successful strategy for mapping agriculture fields in developing countries where data is scarce is to train machine learning models using geospatial labels of fields located in regions with similar climates. The High Plains Aquifer region encompasses Nebraska, Kansas, Colorado, and Texas, spanning a climatological gradient that contains semi-arid and humid regions and supplies nearly one third of all groundwater irrigation applied in the United States. Since the invention of center pivot irrigation in Nebraska in the 1950s, the HPA has experienced significant water level declines as a result of groundwater pumping for irrigation . These were noted as early as the 1980s , yet substantial expansions in irrigated area continued to occur into the 2000’s, with Nebraska adding 1.3 million hectares of irrigated area, a 16.3% increase in total irrigated area for the state . Projections indicate that, assuming pumping at the same linearly extrapolated historic rate, overuse of groundwater will leave 50% of the southern and central HPA dry by 2025 and 2065 respectively . Since the 1980s, management strategies have been evaluated on the basis of whether or not they increase the sustainability of groundwater resources in the HPA while preserving local economies.
These strategies fall into two broad categories: water conservation strategies for increasing water use efficiency or limiting irrigation, and water productivity strategies that focus on increasing yields. Specifically, these strategies have ranged from switching irrigation methods from high pressure applicators to low energy precision agriculture or subsurface drip irrigation , converting land use from water intensive wheat to cotton , converting irrigated farmland to dryland agriculture , and variable rate irrigation. However, found that the dominant crop for drop strategy in Kansas between 1995 and 2005, switching from highly pressurized applicators to low energy precision agriculture systems, still coincided with an unsustainable increase in the total volume of irrigation. The increased profits and water savings generated by the technology switch enabled growers to expand irrigated acreage as well as switch to water intensive corn, alfalfa and soybeans, contributing to a consistent declining trend in the southern and central HPA. Colaizzi et al. 2009 also recounts that for the southern HPA in Texas between 1958 and 2000, despite many center pivot systems switching to more efficient LEPA irrigation methods,vertical grow total irrigation volumes increased with total irrigated acreage. Deines et al. underscores the consequences of a continued present trend of unsustainable groundwater management; they estimate that of the current irrigated acreage across the HPA, possibly 24% would undergo a forced transition to dryland agriculture by 2100, 13% of which is likely unsuitable for crop production, resulting in substantial negative impact to local economies. This body of research suggests that strategies to increase water use efficiency or yields are not by themselves sufficient to sustainably manage finite ground water resources. Groundwater management areas that have the power to tax, educate, and regulate jurisdictions were evaluated on their effectiveness throughout the HPA in Nebraska, Kansas, and Texas. This analysis used well hydrographs, which were missing or absent in some jurisdictions, making a complete analysis of management area effects difficult . This data gap can be supplemented by satellite and machine learning derived estimates of field locations and water use, where irrigation supplied is sourced from groundwater. In regions of the world without good, publicly available records of well hydrographs, remote sensing and machine learning will need to play a role in monitoring and evaluating groundwater use and groundwater management policies. Understanding the spatial location and extent of center pivot agriculture is critical to monitoring, enforcement, and evaluating the effectiveness of groundwater management.
Machine learning models trained on remotely sensed imagery can provide real time predictions on location and extent. In 1986, Strahler et al. identified two classes of models remote sensing – low resolution models where the pixel size is larger than the object of interest and high resolution models where an object is resolved by many pixels . For machine learning problems in remote sensing, the scale of the object of interest determines whether a low resolution or a high resolution model can be employed to locate and classify the object of interest. In many cases, low resolution approaches have been used to classify pixels using only spectral information, even in cases where the high resolution method is available to map objects of interest in finer spatial and semantic detail. Random forest has been the most popular pixel based classifier of choice in remote sensing and has been used to produce global and regional, coarse resolution LULC maps from MODIS imagery . The spatial resolution of the imagery used for global land cover mapping obfuscates any spatial features that can be used to distinguish land cover classes. For global, coarse, land cover mapping, random forest is a more appropriate method than a CNN model pretrained to interpret hierarchies of spatial features. Random Forest has also been applied for regional mapping of particular land cover categories, including irrigated area. While pixel-wise, irrigated area maps are useful, the method used to generate them discards valuable information embedded in the labels, which distinguish between unique fields. Instead of using pixel-wise methods, which only consider spectral information at a point as a feature to classify a pixel, object oriented classification methods that make use of spatial features provided by higher spatial resolution can be used to map objects in finer spatial and semantic detail. These techniques require information about the size, distribution, temporal pattern, and ranges of reflectance in order to be effective, and this can lead to overfitting to particular regions. For circular feature detection, Hough transform methods have been employed, which require prior knowledge of the radii of the objects. However, the Hough transform must be tuned to changes in illumination that happen over a scene or multiple scenes, and also has problems handling changing topographic relief. While it has been successfully used to extract very visually distinct, perfectly circular oil tanks, as well as circular geologic features, each case required visual inspection, manual tuning . In the case of , the scenes tested were very small , allowing for easier manual tuning and less scene diversity for the method to handle. In the case of Cross 1988, there were substantial commission errors. Generally, center pivots are not amenable to this family of techniques since they are not always perfect circles, come in a varying range of sizes which makes the Hough transform more computationally burdensome, can have variable brightness within the field, and are present across the world in varying topographic relief and illumination. All of the aforementioned techniques have a limited capacity to learn complex patterns for object recognition relative to the current state-of-the-art, CNNs. This was demonstrated clearly in 2012 on the task of classifying photographic scenes, with CNNs outperforming the previous state of the art . There is still a need to evaluate how CNN based instance segmentation models perform on coarser resolution sensors like Landsat OLI, given the more limited size of labeled datasets compared to large image datasets like Imagenet and COCO. The simplest model of a neural network can be described as a pipeline which takes as input training data and performs a sequence of linear regressions and nonlinear functions, with each layer in each sequence operating on the output of the previous layer . When used in a supervised machine learning task, the goal in training a neural network is to learn the set of all parameters w and b across all layers of the neural network in order to achieve minimal loss relative to a reference dataset. The learning process is conducted through gradient descent, where a model’s parameters are first initialized and then the model is tested on data that has corresponding reference labels. The difference between the results and the reference labels are then used to adjust model parameters by a learning rate through a method called back propagation, a method for efficient auto differentiation. Model weights are either randomly assigned or taken from a previously trained model, which is referred to as “pretraining” a neural network model.