Interactive map showing crop rotations across Minnesota from 2008-2021
An interactive map viewer to explore crop rotations in the last 14 years in Minnesota and at a variety of scales (whole state, county-level, township-level [~ 1 square mile]). Static thumbnail graphs depicting the crop rotation history for each area are rendered quickly as images on the map to allow initial exploration of trends.
When you zoom the map in or out, the thumbnail graphs and spatial boundaries change to match the zoom level. If you want to explore any given graph in more detail, you can click on it to see an interactive version of the graph with the underlying crop coverage and rotation data, and get dynamic tool-tips when any element of the plot is moused over.
Static graphs of crop rotations appear as icons in an interactive map of Minnesota at the state and county level. If you click on one, a more detailed interactive version of it will be shown to explore the crop rotations for that area in more detail
The app itself is built on top of the MapboxGL Javascript library that provides access to tiled background maps hosted by Mapbox and the ability to add custom interactions with maps.
The map first loads at the whole state-scale, which is its smallest spatial scale it supports. The outline of Minnesota appears along with a static graph showing the state-wide crop rotation data for the crops with the most planted acreage from 2008-2021.
Like any of these static graphs in the app, you can click on it to load an interactive version in a separate panel that allows you to explore the data in more detail using the dynamic tooltips that give the underlying data values.
When you zoom in on the map enough, county-level data will be shown instead, with county outlines and the static crop rotation graph for each county.
Finally, when you zoom in even further, township-level (~ 1 square mile) crop rotation graphs appear.
Crop cover data from USDA is converted to crop rotations from 2008-2021, then aggregated to different spatial scales. I then built custom graphs to visualize these crop transitions more easily.
I first needed to process the annual crop raster data to provide crop rotation information at different spatial scales before I could build any of the web mapping or other features of the app. The USDA's National Agricultural Statistics Service (NASS) publishes detailed crop maps annually based on remote sensing data, and I started by downloading the annual crop raster data for Minnesota from 2008-2021, when the resolution of the maps was consistently 30 x 30m pixels.
I used custom Python scripts to import pairs of the raster datasets from consecutive years and to compare the datasets to tabulate crop rotation information for each pixel in the state over those 14 years.
This information was then aggregated at the township, county, and state levels to calculate the percentage of the given area that was covered by each of the major crops in each year. This also includes linked information about which crops were planted in the same location the year before, and which crops were going to be planted in the same location in the following year.
After tabulating the crop rotations, I then did additional data analysis using custom R scripts. I started by identifying the most widely planted crops state-wide, so I would only include those in the visualizations for clarity. I found that from 2008-2021, six crops averaged over 100,000 acres of identified area (in decreasing order): corn, soy, spring wheat, hay, sugarbeets, and dry beans.
While I was initially planning to only highlight crop rotations to and from these most planted crops, I realized that there was quite a lot of land entering and leaving cultivation each year, most of which cycled to or from identifiable hay crops, so I also included an 'Other' category that represented all other land uses.
I then used R to create the static crop rotation plots for each of the spatial scales, as well as to package the underlying data for each graph into a JSON file that is used to generate the interactive version of each plot when a user clicks it on the map.
To easily organize these graph images and underlying JSON data files, I took inspiration from how raster tiles for mapping (such as with the Leaflet library) are stored and served using a nested folder structure instead of a formal database. The folder hierarchy is standardized to allow easy retrieval of the graph image or the JSON data for any location in the state and at any of the three supported spatial scales.
When a user clicks on any static graph image, the corresponding JSON data is loaded from a URL and an interactive graph is generated from it using the Plotly JavaScript library with the same conventions and graph style as for the static graph. The interactive graphs have dynamic tooltips to access the percentage of area represented by each crop transition (lines) and the percentage of area represented by each of the seven main land uses each year (markers). These plots can be zoomed in or out to see more detail.
I am fascinated by crop rotations, especially to manage plant disease and increase biodiversity. I knew that agriculture in Minnesota varies widely across the state, and I wanted see the different types of crop rotations being used.
After taking a remote sensing class, I was especially interested in how the spectral signatures of different crop plants could be used to map crop locations and acreage for a growing season. I wanted to use the USDA's NASS crop raster data to visualize yearly crop rotations on the landscape in Minnesota.
I chose Minnesota because I knew that farming takes different forms in different parts of the state: while the southern part of the state is dominated by corn/soy crop rotations, the western part grows more sugarbeets and wheat, and the northern part of the state has very little agriculture. I thought these geographic trends could make it an especially interesting state to look at for landscape-level crop rotation patterns.
I built several proof-of-concepts to test major required functions and determine project viability, then I increased complexity from there. This required work in Python and R before the web development.
I knew that I wanted to visualize the crop rotations in a compact, easy to understand way for each spatial area. I also wanted those plots to appear quickly on the map as static image markers and then be shown interactively in more detail when you click on any one of them in the map.
I went through several iterations of graph types to show the data in a compact, but not overwhelming way. I started with trying polar plots, with different angles for the years and variable sized dots at different radii to represent the acreage of each crop planted each year. This became too cramped though, especially for crops represented near the center of the graph.
I then experimented with Sankey-like graphs that show linear flows, in this case showing the flow of planted acreage from each crop to each other crop in the next year. The problem here was that there were too many overlapping 'flows' - 49 of them for each year (connecting each of the 7 land uses to all others).
To try and address these shortcomings, I made three main modifications:
Even though the static and the interactive graphs are made using different tools, they share the same design and colors to make interpretation of the results across the graph types as seamless as possible.
The overall graph format, with static graphs made in R and an interactive implementation of the same graph using Plotly when a user clicks on it.
I'm happy with how the graph format and the interactive version of the graphs together make it as easy as possible to explore trends and changes in crop rotations between years, and across different locations. Being able to get a quick overview with the static graph images on the map and then a more detailed interactive version of the same graph when you click on the static one, I think represents a good trade-off between data density and quickly exploring overall trends by moving the map.
Adding the static graphs to the map required loading them separately and then associating those images with the mapped data. Also, by converting numeric crop codes to strings with padding, I was able to compactly encode crop rotation data and this was key to computing the crop rotations on a per-pixel level.
One of the hardest challenges was to figure out how to have the static graph images appear as symbols in the appropriate map locations. It is not currently possible with Mapbox to provide a URL for an image to use as a marker for each location that's contained within the GeoJSON file that Mapbox uses for plotting locations.
Instead, the fix was to load all images independently from a separate JSON file of image information that I created for this purpose, and then uniquely associating the images with each marker location in the map using matching identifying information in the loaded image and the map locations.
Another central challenge to this project was how to reliably encode crop rotation data between pairs of the yearly crop raster datasets on a pixel-by-pixel basis.
I knew I needed to encode this information using as little information as possible due to the already large data sizes, and in a format that would allow reliable decoding later. In the NASS data, each crop already has a numeric value assigned to it, which I used as the basis for the data encoding.
To compactly track the crop from the previous year and the crop for the next year for each pixel, I used string concatenation of the numeric crop code values for both years (where the crop code for the second year was padded to three characters with leading zeros as needed), and back-converted to a number. This allowed me to reliably parse the three digit crop code for each year from the combined output, starting from the right-most side of the combined crop code.
The table below provides an example of this. During this process, my computer still ran out of memory to complete the computations for the full state, so I split each input file into quarters then ran the computation on each of those before stitching the results back together to get the output for the whole state.
| Year 1 code | Year 2 code | Combined code |
|---|---|---|
| 1 | 1 | 1001 |
| 10 | 10 | 10010 |
| 1 | 100 | 1100 |
| 100 | 1 | 100001 |
This gave me experience tabulating metrics from large spatial raster datasets and mapping the results, as well as experience using custom icon images in Mapbox and creating complex interactive graphs with Plotly.
This project gave me more experience working with Mapbox to display large datasets using custom settings, especially unique custom image markers, which will be useful for future projects. I also gained a lot of experience pre-processing spatial raster data with Python to provide the underlying datasets for use in maps, experience designing clear data visualization, and experience generating custom formatted interactive plots using JavaScript libraries.