Identifying Irrigated Agricultural Land of Alabama Using Remote Sensing and Machine Learning

Over the last century, western states of the U.S. have taken the lead in agriculture. Despite their bountiful yields, their means of production continually drain the available water in the arid climate. However, across the country, Alabama receives more rainfall than nearly anywhere else in the continental U.S., but consistently falls short of the average yearly yield. Where is all of that moisture going? The heavy rainfall is the source for hundreds of streams and rivers that flow throughout the state; however, only a very small percentage of that water finds its way into the fields due to the very limited irrigation.

An initiative started by academic institutions in the state, including the University of Alabama in Huntsville, has worked to encourage farmers to adopt irrigation as an economic and conservation measure. To understand how irrigation may expand, we first have to understand where irrigation currently exists to access the current use of water. To do that, we turn to the Moderate Resolution Imaging Spectroradiometer (MODIS) mounted on the Terra satellite. For this project we are particularly interested in the Enhanced Vegetation Index (EVI). The EVI is calculated using readings in the near infrared, red, and blue part of the electromagnetic spectrum. Healthy vegetation will have high EVI, whereas plants that are stressed or dying will have a lower EVI. We assume that the irrigated crops will have higher EVI readings than the non-irrigated crops.

We use the USDA’s National Agricultural Imaging Program to first identify the agricultural crop land within Alabama’s Houston and Limestone counties. These two counties were chosen because they represent the Southernmost and Northernmost regions of the state. Using this crop mask, we extract all relevant EVI data within this agricultural land. Center pivots are the most prevalent form of irrigation in the state, so we assume that everywhere there is a center-pivot, it is most likely to be irrigated crop land. Thus, in addition to separating the crop land from non-agricultural land, we also use the in-house center pivot survey data to further classify this EVI data into irrigated land EVI values and rainfed land EVI values. This is where our machine learning model comes in.

Our Machine Learning model is first trained on all of irrigated and rainfed (not irrigated) crop land EVI data within Limestone county. Subsequently, this model is tested on the EVI data values from Houston county to predict the ground truth. Then, we use the actual ground truth data to evaluate how the model performed. The machine learning models we use are binary classification algorithms like logistic regression and random forest.

Currently, the bulk of the project is defining a workflow. When we are relatively confident in the results of the current model and the measurements provided by the EVI, we plan on expanding our study to other vegetation indices and satellite products as well as explore the possibility of identifying the varieties of crops and differentiating crop land from surrounding forest and residential areas. The ultimate end goal is to produce a product that can identify the irrigated cropland across the state and then provide estimates of water usage. Once we understand water use, we can begin considering how much more can be used for irrigation in each region while maintaining the health of the river, streams, and groundwater.