Structure from motion matches distinctive features between overlapping images, a step that rests on descriptors such as those of Lowe (2004), and solves at once for camera positions, camera calibration and a sparse set of three-dimensional points; dense matching then fills in the surface. James and Robson (2012) showed that, with adequate imagery, the precision achievable is a small fraction of the viewing distance, which at typical flying heights means a few centimetres. The important errors are systematic. Blocks of parallel, vertical images with a self-calibrated camera tend to produce a broad doming of the model, which James and Robson (2014) traced to error in the lens distortion estimate and showed could be reduced by adding convergent, oblique images. The number and placement of ground control points also governs accuracy, with control needed around the edge of the block and through its height range (Sanz-Ablanedo et al., 2018).

Subtracting one gridded elevation model from another works on gentle terrain and fails on bench faces, where a small horizontal shift produces a large apparent change in height. The M3C2 algorithm of Lague et al. (2013) avoids gridding. It measures the distance between two clouds along the local surface normal, averages over a small cylinder to reduce noise, and attaches to every distance a confidence interval built from local roughness and registration error. The idea that change should be reported only above a level of detection derived from the errors of both surveys is standard in geomorphology (Wheaton et al., 2010), and James et al. (2017) extended it to photogrammetry by deriving maps of point precision from repeated bundle adjustments. Normal distances do not reveal movement parallel to the surface; for that, features can be tracked between orthomosaics by image correlation, as Lucieer et al. (2014) did for a landslide.

Slopes approaching failure commonly show accelerating creep, and Fukuzono (1985) observed that the inverse of velocity plotted against time trends towards zero near the time of failure. Rose and Hungr (2007) applied the method to open pit failures and discussed its pitfalls, and later work has given guidance on smoothing the data and setting alarm thresholds (Carlà et al., 2017). These analyses were developed for prism and radar records sampled many times a day. A weekly photogrammetric survey supplies far fewer points, and that sets firm limits on what it should be asked to do.

The geotechnical engineer names the benches of concern, the failure mechanisms expected from the structural model, and the smallest displacement that would change a decision. That figure is compared with what photogrammetry can deliver at a practicable flying height. If the movement of interest is of the order of millimetres per week, the method is unsuitable and should not be used; if it is centimetres to decimetres, as is usual for bench-scale instabilities in their progressive stage, it is a candidate. Recording this reasoning protects against later over-reading of the data.

Permanent ground control targets are installed around and through the area, on ground judged stable and outside the influence of the benches being watched, and surveyed with the mine's total station or satellite positioning network. They should surround the block and span its vertical range, since pit walls have far more relief than the terrain on which most accuracy studies were done (Sanz-Ablanedo et al., 2018). A separate set of targets is held back as independent check points and never used in the adjustment. Aircraft with real-time or post-processed kinematic positioning reduce the number of targets needed but do not remove the need for checks. In addition, several patches of bare, stable rock are designated as reference surfaces for Step 5.

A single mission is planned and stored: flight lines, height above the wall, camera settings, forward and side overlap, and a set of oblique passes facing the benches. The oblique images serve two purposes. They image the steep faces that vertical photographs barely see, and they strengthen the camera calibration against doming (James and Robson, 2014). The mission is flown at the same time of day each week so that shadows fall alike. Repeating the same geometry does not remove systematic error, but it makes much of it common to every epoch, so that it cancels when epochs are differenced.

Each week's images are processed in the same software version with the same parameters. Images that are blurred or badly exposed are removed first. After bundle adjustment the residuals on the check points are examined, and a survey whose residuals exceed a tolerance set in advance is reflown, not repaired. The products are a dense point cloud, an orthomosaic and a processing report, all archived with the raw images so that any epoch can be reprocessed later. Reporting follows the recommendations for structure-from-motion surveys made by James et al. (2019).

Even with good control, two epochs will differ by a small rigid shift and tilt. Each new cloud is therefore fine-aligned to the baseline using only the reference patches, by an iterative closest point adjustment. The mean and spread of the distances remaining on those patches after alignment become the registration error carried into the next step. If a reference patch itself shows a consistent offset over several weeks, it is no longer stable and is replaced.

M3C2 distances are computed between the baseline and the current epoch, and between consecutive epochs, with normals estimated at a scale larger than the surface roughness of a blasted face (Lague et al., 2013). Distances that fall inside the confidence interval are masked. Areas that have been mined, dumped on or scaled since the previous flight are also masked, using the short-term mine plan, because excavation is change but not deformation. Horizontal displacement vectors are obtained by correlating the two orthomosaics over the unmasked area. Agreement between the two measures, a face moving outward and a crest moving in the same direction, is itself evidence that the signal is real.

The wall is divided into zones that follow the structural domains, and for each zone the median displacement is added to a running series. Cumulative displacement and weekly rate are plotted with their uncertainty. The series are reviewed against the trigger levels in the mine's trigger action response plan. The response this article proposes to a breach is modest by design: inspect the bench, and move a higher-frequency instrument onto it. Weekly photogrammetry is a means of deciding where to look more closely, and its results should always be compared with any prism or radar data covering the same ground (Dick et al., 2015).