-
Notifications
You must be signed in to change notification settings - Fork 0
/
README.Rmd
150 lines (113 loc) · 9.84 KB
/
README.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
---
title: "Cattle Brand Data Pipeline"
author: "Mason Youngblood"
output: github_document
---
The associated preregistration document can be found on [OSF](https://osf.io/d4qv3/), and a detailed description of the model in which this data will be used can be found on the [corresponding GitHub](https://github.com/masonyoungblood/CattleBrandABM).
The Kansas cattle brand books, available from the [state library](https://cdm16884.contentdm.oclc.org/digital/collection/p16884coll5/id/317/rec/2), log all of the brands registered in the state since 1941. Note that Kansas also releases [supplements](https://cdm16884.contentdm.oclc.org/digital/collection/p16884coll5/id/318) between brand books, which log new and transferred brands from the corresponding year. Starting in 1990, Kansas began to use a coding system to catalog all of their cattle brands. Below are the brand books that include this coding system:
* 1990 (9th edition) [scan]
* 1998 (10th edition) [scan]
* 2003 (11th edition) [scan]
* 2008 (12th edition) [PDF]
* 2014 (13th edition) [PDF]
* 2015 (14th edition) [PDF]
* 2016 (15th edition) [PDF]
This coding system includes information on the components that make up each brand (i.e. letters, numbers, symbols), their angle of rotation, and the location of the brand on the animal. Unfortunately, after 2016 the state moved to an [online database](https://www.kellysolutions.com/KS/Brands/searchbrands.asp) that does not include the cattle brand codes. Below is an example of the brand book format (page 60 of 2016).
<center>
{width=100%}
</center>
Each brand code is composed of 13 digits: four sets of three digits that correspond to the components of the brand, and a final digit that indicates the location of the brand on the animal (see below). Some components are a single digit (e.g. "D" and "S" below), while others are two (e.g. "A3" is the caret below) or three. For one- and two-digit components, the third digit corresponds to the angle of rotation of the symbol. When multiple brands have the same set of components in different relative positions, the 12th digit is used to distinguish between the different versions. A detailed description of the coding system, as well as an index of all possible components of brands, can be found in the introduction of each of the cattle brand books.
<center>
{width=100%}
</center>
1990, 1998, and 2003 are scans of the printed books, and thus have noise and artifacts that interfere with optical character recognition (OCR). Manual extraction is extremely labor intensive, so we chose to only collect the cattle brands and zip codes from 1990. 2008, 2014, 2015, and 2016, on the other hand, are the original PDF files. We chose to automatically extract the cattle brands and zip codes from these latter four books using [Tesseract OCR](https://tesseract-ocr.github.io/) ([version 4.1.1](https://github.com/tesseract-ocr/tesseract/releases/tag/4.1.1), compiled [manually](https://tesseract-ocr.github.io/tessdoc/Compiling.html#macos)).
Two custom long short-term memory (LSTM) neural networks were separately [trained](https://tesseract-ocr.github.io/tessdoc/tess4/TrainingTesseract-4.00.html) to recognize (1) cattle brand codes and (2) zip codes:
1. The training dataset for the cattle brand codes included 3,134 currently registered brands (via personal correspondence with [Kansas' Brands Program](https://agriculture.ks.gov/divisions-programs/division-of-animal-health/brands-program)) as well as 500,000 simulated brands based on the possible components outlined in the brand books. Training data was rendered in Courier with the following exposures: -2, -1, 0, 1, 2. The base English-language LSTM was trained for an additional 5,000 iterations using this training data. To further improve accuracy we provided this model with a dictionary containing all of the possible brand components, a set of regular expressions matching the full structure of possible brand codes, and we whitelisted only the characters included in the index of each brand book.
2. The training dataset for the zip codes included all 757 zip codes in Kansas with the state and city, in the same format as in the brand books (e.g. "Bonner Springs, KS 66012"). Training data was rendered in Helvetica with the following exposures: 0, 1, 2, 3. The base English-language LSTM was trained for an additional 5,000 iterations using this training data. To further improve accuracy we provided this model with a dictionary containing all of the zip codes in Kansas and abbreviations for all US states and corresponding regular expressions.
Prior to OCR, all images were manually removed from the brand books using Adobe Acrobat, and all pages were converted to 300 DPI PNG files (same resolution as the training data). OCR was conducted twice for each page, using both the cattle brand LSTM and the zip code LSTM. A series of regular expressions were then used to extract the cattle brands and zip codes from the raw OCR output. Pages that did not result in 57 brands and zip codes (19 rows x 3 columns) were manually corrected. Most errors in OCR were caused by mis-specified brand codes that had an extra component or special characters not included in the index of the brand book. After all data was compiled for each year, we manually corrected any zip codes that did not match a real zip code from a US state or territory.
A random subset of 1,500 brand codes (exported as *brands_to_check.csv*) were manually checked by three research assistants to estimate the accuracy of the OCR pipeline. The result of OCR is a combined dataframe of cattle brands and zip codes with the corresponding pages and years (example below).
```{r}
#load brand data
brand_data <- read.csv("brand_data.csv")[, -1]
#print first 20 rows
brand_data[1:20,]
```
```{r, echo = FALSE, eval = FALSE}
#get subset of 1500 brand codes
set.seed(12345)
brands_to_check <- brand_data[sample(nrow(brand_data), 1500, replace = FALSE), ]
#reorder first by year and then by page, to make checking easier
brands_to_check <- brands_to_check[with(brands_to_check, order(year, page)), ]
#add two columns of 0 for research assistants to replace with 1 when information is incorrect
brands_to_check$error_brand <- rep(0, nrow(brands_to_check))
brands_to_check$error_location <- rep(0, nrow(brands_to_check))
#subset into three objects, one for each research assistant
brands_to_check_elia <- brands_to_check[1:500,]
brands_to_check_lilia <- brands_to_check[501:1000,]
brands_to_check_agnes <- brands_to_check[1001:1500,]
#remove rownames
rownames(brands_to_check_elia) <- NULL
rownames(brands_to_check_lilia) <- NULL
rownames(brands_to_check_agnes) <- NULL
#save
write.csv(brands_to_check_elia, file = "brands_to_check_elia.csv")
write.csv(brands_to_check_lilia, file = "brands_to_check_lilia.csv")
write.csv(brands_to_check_agnes, file = "brands_to_check_agnes.csv")
```
```{r, echo = FALSE}
#check elia
#elia_done <- read.csv("brands_to_check_elia_done.csv")
#sum(elia_done$error_brand) #0 marked incorrect brand codes
#sum(elia_done$error_location) #15 marked incorrect brand locations
#elia_done[which(elia_done$error_location == 1), ] #manual check, actually incorrect are 187, 188, 229
#0/500 actually incorrect brand codes
#3/500 actually incorrect brand locations
#check agnes
#agnes_done <- read.csv("brands_to_check_agnes_done.csv")
#sum(agnes_done$error_brand) #2 marked incorrect brand codes
#agnes_done[which(agnes_done$error_brand == 1), ] #manual check, none are actually incorrect
#sum(agnes_done$error_location) #8 marked incorrect brand locations
#agnes_done[which(agnes_done$error_location == 1), ] #manual check, none are actually incorrect
#0/500 incorrect brand codes
#0/500 incorrect brand locations
#check lilia
#lilia_done <- read.csv2("brands_to_check_lilia_done.csv")
#sum(lilia_done$error_brand) #0 marked incorrect brand codes
#sum(lilia_done$error_location) #6 marked incorrect brand locations
#lilia_done[which(lilia_done$error_location == 1), ] #manual check actually incorrect is 404
#compute overall code and location error
code_error <- 0/1000
location_error <- 4/1000
```
Based on the manual check, we estimate that there was `r code_error*100`% error in brand code transcription and `r location_error*100`% error in location transcription.
```{r}
#load zip code data
library(zipcodeR)
zip_codes <- zipcodeR::zip_code_db
#store unique brands
unique_brand_data <- brand_data[-which(duplicated(brand_data$brand)), ]
```
There are `r scales::comma(nrow(unique_brand_data))` unique brands in the dataset (not accounting for the position of the brands on the animals, so the final number will be smaller).
```{r}
#states of brands
states <- sort(table(zip_codes$state[match(unique_brand_data$location, zip_codes$zipcode)]), decreasing = TRUE)
states
```
`r scales::percent(states[1]/sum(states), 0.1)` of brands are registered in the state of Kansas. Out-of-state brands will be excluded from the rest of the analysis. Below is the distribution of brands within the state of Kansas.
```{r fig.align = "center"}
#get zip code data for plotting
library(choroplethrZip)
data("zip.regions")
kansas_zips <- unique(zip.regions$region[which(zip.regions$state.name == "kansas")])
#get number of brands in each zip code
zip_data <- data.frame(region = kansas_zips, value = sapply(1:length(kansas_zips), function(x){length(which(unique_brand_data$location == kansas_zips[x]))}))
#plot heat map
par(mar = c(0, 0, 0, 0))
zip_choropleth(zip_data, num_colors = 1, state_zoom = "kansas")
```
This distribution is roughly consistent with the most recent cattle inventory of the state (shown below). The most important thing to note is that brand registrations do not cluster around the highest (human) population density parts of the state such as Topeka and Kansas City (in the northeast), which would suggest that they are not registered where they are actively used by ranchers.
<center>
{width=75%}
</center>