Prevalence of the terrain reversal effect in satellite imagery

Figures & data

Figure 1. (a) Pacoima Reservoir, Los Angeles National Forest, as seen on Google Maps. Rivers appear to follow ridges and the reservoir lake in the middle appears as a crater lake. (b) 180° rotation of the same image (without any other alterations) can remove the illusion. Now rivers are in the valleys and the reservoir lake ‘makes more sense’. (c) A negative of the image (again, without further alteration) also removes the illusion (not rotated). (d) The reference terrain represents the real landforms.

Figure 2. Studied locations (where the screenshots were produced). Not all of the 36 locations are indicated in this illustration, as some of them would occlude each other.

Figure 3. (a) and (b) are example stimuli from the first questionnaire, and (c) and (d) are examples from the second. Example questions for these stimuli are: (a) and (c) Point A is located in a place higher/lower than B (b) and (d) Line AB looks like a valley/ridge.

Figure 4. Basic experimental design. NH_R: Northern Hemisphere rotated 180°, SH_R: Southern Hemisphere rotated 180°.

Table 1. Participant information in the experiments

	Experiment 1		Experiment 2
Participant information	Number	%	Number	%
Age
<19	0	0	76	27.5
20–29	60	23.9	31	11.2
30–39	115	45.8	52	18.8
40–49	44	17.5	62	22.5
50–59	24	9.6	40	14.5
>60	8	3.2	15	5.5
	253	100	282	100
Education
University	239	94.8	185	65.6
No university	14	5.2	97	34.4
	253	100	282	100
Specialization
Trained in GI	189	75	94	33.3
Not trained in GI	63	25	188	66.7
	253	100	282	100
Gender
Male	178	70.6	164	58.1
Female	74	29.4	118	41.9
	253	100	282	100

GI: Geographic Information Science.

Total number of participants in the two experiments: 535.

Figure 5. Examples from four different online map providers demonstrating that the terrain reversal effect can be found in many of the current services (web services accessed in June 2014).

Figure 6. Interaction plot showing how the percentage of correct answers changes over all tasks in the two experiments. NH and SH images have clearly opposite success rates and after the rotation, SH images lead to more incorrect answers while NH images lead to more correct answers. The three images from the equator do not show a similarly strong tendency.

Figure 7. Box-and-whisker plots of raw data showing percentage of correct answers per stimuli group. The solid line inside the box is the median. Boxes cover the interquartile range, whiskers extend 1.5 interquartile range from the boxes, and outliers outside this range are indicated with points.

Table 2. Estimates of fixed-effect parameters of binary generalized linear mixed-effect models on questionnaire data.

	Estimate	Standard error	z value	p value
Intercept (NH)	−0.33	0.17	−1.90	0.057
SH	1.68	0.24	6.87	6.61 × 10^−12
NHR	1.60	0.07	21.90	<2 × 10^−16
SHR	−4.58	0.08	−55.82	<2 × 10^−16

Figure 8. Estimated correct answer rates based on inferential statistics for images from NH and SH before and after the rotation (NH_R and SH_R). The values were obtained from the generalized linear model estimates presented in Table 2 and transformed back to the response scores (expected success rate).

Table 3. Estimates of full model

	Estimate	Standard error	t value	Pr(>∣t∣)
(Intercept)	0.2283	0.1457	1.57	0.1173
Age	0.0001	0.0001	0.60	0.5499
Gender:male	0.0416	0.0209	1.99	0.0471
Education	0.0036	0.0264	0.14	0.8902
Geopro	0.0315	0.0156	2.02	0.0437
ThreeD	0.0099	0.0154	0.64	0.5206
Photoconf	0.0032	0.0261	0.12	0.9010
Colorvision	0.0426	0.0497	0.86	0.3908

Note: Except gender, all other variables take numeric values. R² = 0.02, Adj-R² = 0.01, F(7,940) = 2.847, p value = 0.006.

Table 4. Estimates of the model with gender and geoprofession as predictors.

	Estimate	Standard error	t value	Pr(>∣t∣)
(Intercept)	0.3815	0.0727	5.24	<0.0001
Gender:male	−0.0214	0.0932	−0.23	0.8181
Geopro	0.0290	0.0172	1.69	0.0918
Gender:male × Geopro	0.0149	0.0217	0.69	0.4927

Note: R² = 0.02, Adj-R² = 0.02, F(3,944) = 6.305, p value = 0.0003.