Plants of Southern California: Accuracy of iNaturalist Positions


Table of Contents

Summary
Introduction
Data Analysis
Sources of Large Position Errors
Tips for Producing Better Positions


Summary

iNat positions are generally not as accurate as GPS positions made from a continuously-running handheld GPS unit. 50% of iNat positions can be relocated within 40 feet; 90% can be relocated within 100 feet; and 10% can have an error over 100 feet, with a very small number with errors as large as ~10 miles. The error for the 90% relocated within 100 feet is primarily caused by GPS initialization error, which happens for every set of photographs because smartphones and GPS-enabled cameras shut off their GPS processors immediately after a position fix is obtained.

Large errors are mostly due to observers not using a GPS source and mislocating their observations, but are also due to smartphones and GPS units experiencing a problem in getting a good signal, including poor satellite geometry; shielding of satellites by the terrain; multipath errors due to reflection off boulders; and infrequent glitches in GPS units.

Introduction

iNaturalist is a wonderful platform to share plant locations, and I and others not-infrequently try to follow-up some observations for various reasons. The most exciting reason is to see a species that we've never seen before, and have wanted to see. For those of us working on the flora of an area, iNaturalist can alert us to the presence of species not yet recorded in our floral areas. I also sometimes followup iNat observations in a region I've done a flora of, in order to check on their determinations, when the determination is not clear from the photographs alone.

One major iNat benefit is to add more accurately-located points to species maps. Voucher maps of a small area are often wildly imprecise, since many vouchers have only vague locations ("San Jacinto Mountain"; "west edge of Sonoran Desert"), yet have to be given coordinates as part of georeferencing all vouchers.

For all of the above purposes, accurate locations are essential. I naively originally assumed that because most iNat locations are taken with a smartphone, or a GPS-enabled camera, that those locations would be as accurate as ones taken with a hand-held GPS unit. Although that turns out to be not terribly-incorrect for perhaps half of iNat locations, it is definitely not correct for the other half.

I and others have been not-infrequently-puzzled by not being able to find the plant photographed in an iNat observation at the coordinates of that observation within the typical 20 to 30 foot radius expected when trying to relocate a GPS point taken with a hand-held unit. Sometimes it is clear that an observation had to have been taken miles away from the iNat location! When a friend told me that he tried to follow up five separate observations that had been posted over a period of one month, and that he couldn't relocate any of them, I decided to start trying to relocate a few plants on each of my hikes, and recording the relocation errors in a database for analysis.

Data Analysis

Over the last year, I have recorded 23 data points through making repeat observations of an iNat observation with my GPS-enabled camera (an Olympus TG-6), in a variety of settings from the desert to the mountains. I attempted to locate another iNat observation, but I couldn't relocate it despite attempts on two different field days.

Ordinarily, I would not have published an analysis of position accuracy based on just 23 data points, and waited until I had a larger data set. However, I accidentally discovered the source of the larger errors for most iNat locations in what I thought was a completely-unrelated investigation of a larger data set, so this webpage presents the results from both data sets.

The unrelated data set came from 75 iNat observations that I posted from a day on which I stayed completely on a trail. I looked at the iNat map of those observations just to make sure there weren't any glitches in the positions, and I was shocked to see that 21 of those observations did not plot on the trail! That never happens when I look at a map showing GPS points taken on a trail with a hand-held GPS unit; it is rare that any GPS point out of a hundred or so has any noticeable error looking at a similar map.

I derived position errors of those 75 camera observations by comparing the camera location with the closest GPS location in time with a continuously-running Garmin Etrex 10 hand held unit set for maximum accuracy (WAAS off; GLONASS off; recording interval "more often"). Both the camera and the GPS received their time signal from the GPS satellites, so their time was synched. I checked the track from the GPS, and it matched the trail essentially perfectly.

An analysis of the position errors of those 75 observations gave results similar to the distribution of the iNat errors based on those 23 points. The cumulative distributions of the errors from those two sources are presented in Fig. 1.

Fig. 1. The cumulative distribution of the number of observations within a given radial error comparing my camera GPS location to the iNat observation location (except for one iNat location that couldn't be relocated), and comparing my camera GPS location to a Garmin Etrex 10 unit.

The curves in Fig. 1 are fairly similar. The camera vs. Garmin curve is somewhat worse due to the more challenging terrain in which that comparison was done. It was done on the PCT south of S22 / Barrel Spring, which has poor satellite visibility in many places due to the trail being significantly below a ridge, with a number of incursions into drainages where as much as half the sky is blocked at times. Many of the relocated iNat positions were in the open desert with a good view of the entire sky.

Fig. 1 shows that roughly 50% of all iNat locations, and my camera positions, are within 40 feet of the original iNat location and the Garmin Etrex position (actual numbers are 48% vs. iNat and 57% vs. the Garmin; I used bins of multiples of 20 feet).

40 feet is not a terrible error; it is certainly good enough to relocate a given plant, even a tiny one with some searching, and way better than needed to make good species distribution maps. But only 50% of repeat observations have that error, instead of the ~99% one would expect if a hand-held GPS unit was used to take both the original and repeat observation, assuming the GPS unit had been in operation long enough to avoid initialization errors.

Worse, there is a very long tail for the other 50% of observations. 10% of all iNat relocations have an error more than 100 feet; 10% of my camera / Garmin comparisons have an error of more than 160 feet. The worst iNat repeat error in this set was 626 feet; the worst camera / Garmin comparison error was 458 feet. Those kind of errors were totally unexpected by me for either source. Note that the one iNat observation that couldn't be relocated might or might not have had a larger error.

The source of the long tail is almost surely mostly, but not always, due to GPS initialization error.

When GPS units are first turned on, they frequently have a large initialization error, that decreases the longer the unit is on. This is usually shown by what the unit thinks is the accuracy of the position, which starts at 100 feet or more and then gradually declines to ~10 feet. Even if the unit has a recent fix, if you turn off a GPS unit and then turn it back on, it will still have a relatively large initialization error, since it doesn't know how far you have moved during the time the unit was off. One egregious example of this was, after having my GPS on all day on Black Mountain Road at San Jacinto Mountain, I had turned it off near the bottom of the road, thinking I was done with making observations for the day. But within five minutes, I found something else I wanted to record. When I turned the unit back on, it reported that I was MILES away from my current location, at an elevation several thousand feet below sea level. It would have taken the unit a very long time to walk itself out of that strongly-erroneous position, so I turned it off, then back on, and this time it gave the correct location.

I had naively expected both smartphones and camera GPS units would run their GPS processors continuously, avoiding initialization errors. But they don't, because to do so would drain their batteries faster. As a result, both of them only turn on the GPS when requested by the camera, which means every camera photograph has a GPS location that suffers from initialization error.

If you look closely on a map at the locations of a set of observations posted at iNat made from a smartphone, that has a trail or road also shown, you'll often see something like the map shown in Fig. 2, which shows some observations are "off trail".

Fig. 2. Map of iNat observations (blue diamonds) I posted from a hike on the PCT south of S22 / Barrel Spring on 3 April 2020, and a GPS map of the trail (red line) made from a continuously-running Garmin GPS unit from one trip, and confirmed to the width of the red line from a number of other trips. Note that the trail is mostly on the east side of a fairly-steep ridge, with incursions into drainages where the satellite signals could be blocked from as much as half the sky. The points farthest from the trail are most often found near those drainages.

Note that seven of the observations are obviously quite far away from the trail (over 200 feet away) and another 14 or so are noticeably off the trail (over 80 feet away). In comparison, the track from my continuously-running Garmin on the same trip has every point from both the outgoing and ingoing track within the width of the red line that marks the trail, with one short exception where the outgoing and ingoing tracks are different by up to 100 feet for a very short distance.

Clearly, a devoted GPS unit that is continuously running is able to give much better accuracy than a unit of any kind that is only sporadically turned on for a brief time. However, even devoted GPS units have a hard time in challenging terrain, like parts of the Devils Slide Trail at San Jacinto Mountain (see its note about GPS signals), and how many points are off-trail at the Bright Angel Trail in the Grand Canyon. I had to spend considerable time and effort to get good points for my 2019 observations there, despite a continuously-running GPS unit. The linked plot would have had many more off-trail points if I hadn't hand-corrected my points.

A secondary contributor to the error is that smartphones only take GPS points infrequently, even when the GPS unit is active, and hence the GPS location assigned to a photograph may reflect where the observer was as many as ten seconds previously. Fred Melgert wrote:


I've been working on the BorregoHiking App that is able to record a track. In general Android does everything in its power to prevent frequent GPS recordings. It takes an awful lot of battery resources to get a GPS fix.

On the phone while recording, it's not unusual to get an update only after 10 seconds. A travel distance of 20-40 meters during that time isn't unusual.

The phone simply refuses to update very frequently.


Sources of Large Position Errors

These are all errors that I've noted while either looking at, or trying to relocate, iNat observations. I couldn't include them (with one exception noted below) in the statistical analysis above, since I didn't keep track of the much larger number of iNat observations without large position problems.

Large errors due to observers not using a GPS source and not accurately knowing where they were. Here are some examples of problems that I've seen, with the actual locations generalized, and the pronoun "he" used in all cases, so as not to embarrass the observer. I don't want to single anyone out, since these are generic problems.

Large errors due to GPS problems.

GPS itself can have problems, even with a continuously running handheld unit, due to a problem in getting a good GPS satellite signal. This can come from poor satellite geometry; shielding of satellites by the terrain; multipath errors due to reflection off boulders; and infrequent glitches in GPS units.

Examples:

Tips for Producing Better Positions

The best method of getting good positions is to have a separate continuously-running GPS unit that can be used in different ways. The best solution feeds coordinates from that GPS to your smartphone or GPS-enabled camera, which also then allows you to turn off the GPS in those devices. Another solution is to use its coordinates, at home after you download the data, whenever you notice a problem with your iNat positions posted using the smartphone or camera's positions.

However, most iNat users will not want to buy a separate GPS unit, so the next-best method is to make sure you are getting good positions by optimizing your unit.

For a smartphone, use one or more of the following approaches:

If you are having a problem with GPS, there are webpages to help you, such as one for Android phones and one for iPhones.

If you use the Android iNat app, you can check the "accuracy" in the app (shown under the location name, to the right of the little pin icon, in meters) before you take your photos. However, the iPhone iNat app only shows the accuracy after you make an observation, so you can check it then. If the accuracy is poor, you can try again.

For a GPS-enabled camera, use one or more of the following approaches:

Note that turning on your camera, and leaving it on for a while, is not one of the suggested options to produce better positions, since smartphones and cameras try hard not to turn on the GPS chip unless requested. Some of the tips above are work-arounds to make the smartphone check GPS more often.


I thank Fred Melgert and Don Rideout for discussions that stimulated this analysis and provided some of the information for this page.


Go to:


Copyright © 2020 by Tom Chester
Permission is freely granted to reproduce any or all of this page as long as credit is given to me at this source:
http://tchester.org/plants/analysis/iNat/position_accuracy.html
Comments and feedback: Tom Chester
Last update: 9 April 2020