Fig 1.
Scene familiarity landscapes from satellite images.
A Satellite images (zoom level = 18; ~250 m camera altitude) of an area encompassing the University of Sussex campus in Falmer, England were stitched together in MATLAB to form a single 3,600 x 3,600 pixel grayscale image (scale bar = 100 m). The focal scene (highlighted by the right-most central white circle) is centered on the Meeting House on the Sussex campus (longitude: 50.8649; latitude: -0.0880). Ten thousand 640x640 scenes were sampled at equally spaced points in a 100x100 grid (white dots) across the larger landscape. Each scene was processed to 50 x 50 resolution with 10 levels of gray. These sample scenes were circularized prior to similarity processing. B Expanded views of the central scene and a scene shifted two sampling points to the left. These locations are indicated by white circles in A. C All 10,000 scenes in A were compared to all other scenes. The contour plot indicates average image difference values by scene position. The plot in C is superimposed atop the campus satellite image for reference. Dotted rectangles d and e show regions of relatively high and relatively low distinctness. These two positions are used to generate the volcano plots shown in D and E. These plots show the absolute pixel-by-pixel image differences of the focal scene compared to the surrounding 225 scenes (15 x 15 square). The red arrows below the volcanoes indicate the orientation of the best-matched surrounding scene rotated through 360 degrees and compared to the static focal scene; arrow length varies directly with the goodness of the match. Satellite images used under a CC BY license, with permission from Esri, original copyright 2015.
Fig 2.
Comparator measures of landscape information.
A Slices are made through the familiarity volcano in the four cardinal directions and the values are averaged to produce the curve shown in B. The absolute scene difference value and the distance from the focal point (in meters) are derived from this curve at the 50% maximal difference point (p50).
Fig 3.
Schematic of route tracking algorithm.
The yellow dotted arrows indicate a user-selected route. The blue arc scans are based on the direction of the most familiar scene at that point (dotted black lines). Initial samples are made in the direction of the current path with subsequent samples taken at progressively wider angles on the arc, alternating left and right views (green points). At each point sampled, the agent rotates the focal current scene through 360 degrees and compares each rotation to all stored memory scenes (yellow path). Once a familiarity threshold is met, no further points on the arc are sampled. The agent moves forward to this best-matched point (red line) and casts its next scan based on an extension of this red line segment (see Methods for further description).
Fig 4.
The spatial information available in natural scenes.
A Five sampled focal scenes are shown as white dots superimposed on satellite regions for Derwentwater, Mojave, and Amsterdam (scale bar = 100 m). B The averaged volcano plots for these five scenes, each comparing focal scenes to the surrounding 400 scenes (20 x 20 grid) are shown for a 20 x 20 pixel-resolution with 100 gray level pixel depth. A common y-axis is used to show information differences among regions. C RIDF curves averaged across the five landscape sample points for the four sensor resolutions indicated in the legend (all at 100 gray level pixel depth). Absolute scene difference information (D) and distance from focal scene (E) are plotted for each region for each of the sensor resolutions and pixel depths as calculated at the p50 position on the information volcanoes (as described in Fig 2). Satellite images used under a CC BY license, with permission from Esri, original copyright 2015.
Fig 5.
Route recapitulation performance across different landscapes.
We tested the ability of the auto-tracking algorithm to retrace simple, S-shaped training routes (yellow dots) drawn across the Derwentwater, Mojave, and Amsterdam satellite regions (scale bar = 100 m). A Sample retraced paths are shown for two sensor resolutions (10 x 10 and 40 x 40) and 10 gray level pixel depth (insets show examples of sensor images). B Cumulative distance of departure of the retraced route from the training route is shown for the three regions at the four sensor resolutions and three pixel depths. Plotted are the summed geometric distances of each best-matched point (red) to the appropriate training path point (in sequence) based on a 100 x 100 point sampling grid imposed on each region (actual values given above bars). Satellite images used under a CC BY license, with permission from Esri, original copyright 2015.
Fig 6.
Examples of tracking algorithm performance on complex routes.
A Retracing of a route in the shape of the University of Sussex logo over the Sussex landscape. This route had several abrupt turns, including one of about 135 degrees. B Test of the tracking algorithm on a route that loops across itself. This route spells out a scripted “Jan” in honor of an uncle who lived in this Amsterdam neighborhood. C An expanding square spiral beginning at the Meeting House on the Sussex campus. The tracker accurately retraced the route until it passed through the region of low scene difference information over the field at the lower left of the image (near region e in Fig 1C; scale bar = 100 m). Satellite images used under a CC BY license, with permission from Esri, original copyright 2015.