Verifying Volunteer Entries to the NYC Street Tree Census

Summary

Using NYC Open Data's 2015 Street Tree Census Data Set (click Export to download a non-geospatial version of the dataset), I conduct exploratory data analysis and feature engineering to find the significance of certain variables on the health of NYC street trees.

I then build a classification model to gain further insights into which variables play a role in a tree's health.

My goal is to determine which features most help my model perform well and what information may be helpful for future censuses and policy in general.

Finally, I generate a list of volunteer data entries for which the health status does not match that of my model's predictions. This list could be used by professionals to double-check the work of volunteers and re-determine a tree's health if necessary.

Objectives

1. Investigate any notable differences between professional and volunteer entries to the census.
2. Explore data to discover features that are important in determining the health of a tree. Use this information to make policy recommendations. Additionally, determine if there are other data that could be collected in the future.
3. Engineer features to improve model and see if any data should be added in future censuses.
4. Build a classification model using only professional entries to determine the health of a tree. Use it to predict tree health for volunteer entries and compare results with volunteer judgments.
5. Use maps to bolster discoveries, if possible.

Findings

1. Professional entries vs. Volunteer entries

• Volunteers did not make it out to the outer edges of the city and tended to rate trees more poorly (lower proportion of trees in Good health compared to professional staff).

Tree health breakdown for professional and volunteer entries

	Professional	Volunteer
Good	81.8%	79.5%
Fair	14.2%	16.1%
Poor	4.0%	4.4%

2. Discover important features already present in data

• Species of tree is statistically significant.
  • Norway maple is the species with the lowest proportion of trees in Good health, and sawtooth oak is the species with the highest.
  • I would recommend sticking to the healthier varieties (toward the right side of the graph seen below) and avoiding the less healthy varieties (toward the left of the graph).
    • NOTE: graph sorted, in ascending order, by proportion of trees in Good health.

• Problems with the tree itself (i.e. problems present in the roots, trunk, or branches) were obviously significant.
  • Unfortunately, problems that were listed as Other appear to hold the highest significance.
  • For the next census, I recommend including columns with more specificity and/or a notes column (the latter of which one could analyze using NLP).
  • I suggest conducting more regular maintenance, as well as pushing for environmental protections, in the style of the plastic ban bag.

• All levels of neighborhood delineation held significance.
  • I chose community board as my neighborhood variable in part because one must petition the community board to make changes to street trees.

3. Feature engineering

• block_count - the number of trees on the block.
  • I wanted to see if tree density had anything to do with a tree's health, but found no discernible relationship during EDA. My classification model, however, told a different story (see below).
• neighbor_dist - distance (in meters) to the nearest tree.
  • Similar to the above feature, I wondered if two trees being close together would negatively impact their health. Again, during EDA, I found no relationship. And again, my classification model told a different story (see below).

4. Modeling

• It was difficult to choose which model to use based off of numbers alone. I had to think carefully about how it was being used.
  • I went with a model that was more inaccurate but had a better spread of predictions. While it may not be ready for implementation yet, it is a step in the right direction.
  • I created a metric--precision of Good predictions--that served as a fairly good proxy for the desired results, though more investigation may be necessary to find an even better one.

• Many of the most important features--including, but not limited to, the number of trees on the same block, distance to the nearest tree, and presence of sidewalk damage--didn’t show much significance when looked at during EDA, but must have had solid predictive power after interacting with other branches of the Random Forest trees.

Top features in the model (in order of the average decrease in Gini impurity), out of 147 total features:

Distance to nearest tree    (15.9%)
Tree diameter               (14.0%)
Number of trees on block    (12.9%)
Species [Norway maple]      (3.0%)
Trunk problems [other]      (2.8%)
Branch problems [other]     (2.7%)
Tree stewards [1-2]         (2.0%)
Sidewalk damage             (1.8%)
Root problems [stone]       (1.4%)
Species [London planetree]  (1.3%)
Species [Honeylocust]       (1.3%)
Branch problems [light]     (1.2%)
On curb                     (1.0%)  
Community board [414]       (0.9%)
Root problems [other]       (0.9%)

5. Maps

• Maps were indeed useful in visualizing areas in need of tree improvement throughout the city, as well as comparing professional and volunteer entries (as shown above).
• To look at and recreate my maps (or build your own), refer to the mapmaking notebook in this repository.

Conclusions

I ran my final model on the data collected by volunteers and compiled a list of trees for which the health status determined by volunteers differs from that predicted by my model. The list totals 78,645 trees, which is a bit higher than is feasible to check in the real world. If a team of ten people each looked at 100 trees a day, it would take them roughly four months to look at every tree.

Still, I was able to use the data from the 2015 census to back up policy proposals that make sense for NYC's Department of Parks and Recreation and are possibly already in the works. Furthermore, I identified some key areas of improvement for data collection during the next street tree census in 2025.

Next steps

I could potentially create a better model using a neural network, but that may be too computationally expensive with a dataset this large.

Instead, I think the next steps for this project will include transforming this into a binary classification problem (i.e. combining trees with Fair and Poor health status to compare trees in Good and Not Good health). This would surely improve the model's metrics and shorten the list of trees to check, while still addressing the project's goals.

In the meantime, the NYC Street Trees Census is fairly thorough and rife with opportunities for data exploration and predictive modeling. In future censuses, even more data could be gathered (especially regarding the specificity of problems with a tree's roots, trunk, and branches) that will improve prediction even further.

List of files

• archives folder - scrap notebooks.
• data folder - cleaned data and variable descriptions.
• images folder - insightful charts and maps from the project.
• .gitignore - list of files and folders to ignore.
• README.md - this very file!
• eda_modeling_evaluation.ipynb - Jupyter Notebook for data cleaning and exploration, feature engineering, and classification modeling.
• functions.py - text file with functions for data cleaning, feature engineering, statistical tests, and visualizations.
• mapmaking.ipynb - Jupyter Notebook for making maps of NYC street trees.
  • NOTE: this is meant to be more of a sandbox than a polished workbook.
• presentation.pdf - presentation for New York City Department of Parks and Recreation with my findings.
• volunteer_verification.ipynb - Jupyter Notebook for using my model to predict on volunteer entries and build a list of trees for which the health status determined by volunteers differs from that predicted by my model.

Check out my blog post