The toolbar's folder means “bring this into the image I am looking at”: one raster becomes a layer over the current image, and several ask whether to layer them or combine them into one all-band image. File ▸ Open makes a new window. Files can also be dragged onto the window, and a file the app is launched with takes the same door as any other.

A satellite scene routinely ships several resolutions over the same ground: Landsat 7 delivers six 30 m reflective bands, two 60 m thermal bands and a 15 m panchromatic band. Those cannot become one raster without resampling, and nothing is resampled on open. The grids are kept separate, and each is described by band count, pixel dimensions, ground pixel size and sample type. Each grid has a tick. A popup over the ticked rows chooses which becomes the window's base. The rest arrive as layers over it in the same window, each at its own native resolution. Open is refused while nothing is ticked.

The folder is walked and every raster it holds is offered as one selection, with the same Layers-or-Combine question a multi-file open asks. There is no limit on how many may be opened. Tiles that sit side by side combine into one image, and a survey folder of 92 orthophotos becomes a single mosaic over the whole extent. The one ceiling is on walking the tree, at 4,096 files. When it is reached, the count is reported as a floor.

Nothing is corrected on load. The default stretch is Original, the band's full range, and no calibration, mask or resampling is applied. The catalogue reports what conversions the delivery offers. Applying one is an explicit act.

Split turns a multi-band stack into one layer per chosen band, for when bands need separate display settings or separate processing. Combine is its inverse. Single-band layers sharing an exact grid become one composite by reference, with no pixel copied. The bands are ordered red, green, blue, near-infrared and then by wavelength, the same rule the import uses. A split-and-recombine round trip lands each band back in its slot.

A window over the imagery you have. It lists every raster in your Workspace folder and any folder you add, with a thumbnail, a name and the facts that decide which scene you want: satellite, level, date, format, grid, bands, CRS, pixel size and size on disk. A delivery is one row read from its manifest. Loose band files of one scene regroup into one row per grid. Rows appear before they are described and the sidebar counts how many have been read. Sort and filter from the toolbar. Double-click a row to open it, through the same preflight a File ▸ Open would give it.

A window of step-by-step rails through the longer tasks: a proximal survey to soil pits, a property's carbon-forestry baseline, change between two dates, an unsupervised land-cover map, a DEM to a draft Land Resource Inventory, raw numbers to comparable reflectance, taking sensor striping out, several scenes into one mosaic, sharpening coarse bands with a finer guide, and finding and downloading the right scene. A shelf of cards shows each rail with a miniature of its progress. Open one and the steps run along a traverse, with the current station ringed. Every step names the menu its controls are under.

Every raster, vector overlay, polygon and annotation layer in one drag-reorderable tree, with groups. Each row carries a visibility toggle, an opacity slider and a blend mode. The base raster is a member of that stack like any other, and a layer can sit below it. Clicking a layer's name makes it the active layer, marked with an amber scope glyph. The active layer is the target for processing, for the Bands panel and for the readout. Right-click for Rename, Remove, Save As GeoTIFF, Copy, Open in New Window, Use as Processing Region, Relief Shading, View in 3D, and the elevation-model statement.

Normal, multiply, screen, overlay, soft light and luminosity, the six ISO 32000 modes, on any raster layer, blended in the same render pass as everything else. A group has its own picker. Pass through leaves the group as a folder whose opacity multiplies into each child. Any mode composites the group's layers into one picture first, then blends that picture over the frame at the group's opacity. Those are different images: white in multiply inside a group at 50 % over grey reads 0.5 as pass-through and 0.75 as a picture.

Categories covering georeferencing (CRS, pixel and ground size in the CRS's own units), sensor identity and the resolved calibration profile, acquisition date and sun geometry drawn on a sky dome, the composite's source file and band for each slot, applied display geometry such as crop and rotation, raw metadata, and a System category showing the GPU, memory gauges and streaming state.

The footer is the instrument panel: cursor coordinate, pixel values, display scale as a percentage where 100 % is 1:1, the scale bar, and an SDR/HDR badge with the display's EDR headroom. The coordinate's decimals follow the pixel size: six over 10 cm imagery and four over a kilometre. Its context menu switches between decimal degrees and degrees-minutes-seconds. Right-clicking the map copies the clicked coordinate as signed decimal degrees, the form other tools accept when pasted. Rulers draw along the top and left edges. A ruler labels one axis along a screen edge, which only holds while the two are parallel, and the rulers switch off while the image is rotated.

Where the scene is, on a basemap. Every raster in the window is drawn as a footprint, and the pane follows the active project. A zoomed-in view still says which coastline, city, river or lake it is over. It draws from a national tile service where one covers the ground (LINZ, USGS, swisstopo, IGN) and falls back to Apple's own basemap where none does. The pane says which it is showing. It also says when it has magnified past a service's deepest tiles.

Shapefiles, GeoPackage, GeoJSON, KML and KMZ open as overlay layers drawn in the raster's own coordinate space, alongside the polygons you draw, imported point layers, GCP markers, classification results and annotation notes. Each is its own layer: hide it, fade it, reorder it, or flag it Use as Processing Region so that region-aware jobs clip to it exactly as they clip to selected polygons.

A Trap.NZ project is drawn over the open image as one group of layers from its map feed: the project boundary, its lines in the colours the project gave them on trap.nz, and its traps coloured by how long since each was last recorded against. Click a trap for its callout: code and line, type, last record with the days, installation, tags, sensor, notes. Shift-click to add another, and click the dot again to put it away. The menu's Show Details for All Traps tick draws all of them for a figure. Unticking it puts every callout away. Retired traps are counted in the banner and not drawn. The bands are the calendar's: within a week, a fortnight, a month, two months, a quarter, half a year, longer, and never recorded. The Overlays inspector draws the legend as a bar per band by count, on the colour strip, over a scale of days. Showing a project again refreshes it in place. Lines and dots arrive at a width made to be seen. The row's Line Width and Point Size sliders change either.

Setup is a key and a list. Generate a map-feed key under My account on trap.nz and paste it into Settings ▸ Trap.NZ. Then add projects by id. The id is the number at the end of the project's address, https://trap.nz/node/123456. Alternatively, let Look Up My Projects list the ones your account belongs to. The key is read-only, kept in your keychain and sent only to io.trap.nz. Because Trap.NZ carries the key in the feed's address, it is never written to a cache.

Draw a line across an elevation model and the pane charts what is under it: distance along the ground on one axis, the surface's own height units on the other. Because a profile has to say which surface it measured, it is offered only when the active layer holds heights.

Drapes the layer over an elevation model in the same window, under an orbit camera, with vertical exaggeration on screen. The DEM is chosen by the most ground it shares with the image. Where the image reaches less than all of the surface, the menu item's help states the share, rounded down. The image lands at its true ground position and scale whatever its bounds are.

Drag a rectangle over the active layer to keep. Shift squares it. The marquee snaps outward to whole pixels of the layer being cut. A marquee dragged around something keeps everything it enclosed. Cropping again intersects with the existing window. It is display geometry. No sample is discarded, and ⌘Z undoes it. It carries through to a GeoTIFF export as a source window, and the written pixels are the file's own.

Drag the dial to turn the active layer to any angle, 0 to 359 degrees clockwise. No pixel is resampled. Rotating the base raster turns document-pixel space itself. Polygons, imported vectors and annotations follow the imagery they were placed on. Info states the offset from the source in the file's own pixels.

The base document is A and any added or derived raster is B. Four modes: fixed side by side, a draggable swipe, a linear A/B blend, and timed flicker. All are display-only and isolate A and B from the rest of the stack. Pressing the glyph again stops comparing.

Clicking the image drops a callout: a dot pinned to that pixel and a text box standing off from it, joined by a leader line. The dot is held in image pixels and stays on its feature through pan and zoom. The box keeps one readable size at every zoom. The distance between them is never a ground measurement. Selecting a callout shows a style bar with five colours: box outline, box fill, line and dot, text, and a highlight behind it.

An on-device assistant for questions about remote sensing, the sensor catalogue and how this app does things: which bands make a natural-colour composite of the scene in front of you, what separates NDVI from NDMI, how to pan-sharpen. It runs on Apple's on-device model, and nothing leaves the Mac. Each answer names where its facts came from.

One lane per platform, a tick for an instant and a bar for a span, the marks placed at true time. Click a mark or drag the playhead. The playhead snaps to observations, and the observations never move. Following the playhead leaves exactly one date checked in the Layers sidebar. Drag across the strip to set a window. A composite runs over that window, and the panel names it. A pixel's series is charted beneath as you move the cursor. Its line breaks at every date that saw nothing, and each gap says which kind of nothing it was: unobserved, masked, not computable, unreadable. The series exports as CSV with the empty rows kept.

Reduces the dates in the window to one raster. A per-band reduction (median, mean, minimum, maximum) takes each band separately. The red may come from March and the near-infrared from July. The result is what the ground usually looks like, and no sensor recorded it. A donor rule (largest or smallest value of an expression you write, earliest, latest) picks one date per pixel and copies its bands whole. The output is still a measurement, with seams where neighbours chose differently.

One row per band. Each row's three dots put that band in the red, green or blue channel, and clicking what the row says the band is lets you correct or supply it. Above the list, presets (True colour, Colour infrared, SWIR, Agriculture, Geology) are each drawn as a small spectrum showing where their three bands sit. A preset whose bands the image lacks draws them hollow, and the reason it cannot apply is visible. Where bands state their wavelengths there is a spectrum strip whose R, G and B markers drag along it, or step band by band with the arrow keys.

Original (full range), percentile clip, manual low/high, and empirical-CDF histogram equalisation, each with gamma. The default when an image opens is Original.

Bins the shown bands over one shared domain so the channels can be compared. Scope switches between the viewport and the whole image. For integer data a bin is a whole number of values wide, and a gap means no pixel holds that value. The bars are drawn as steps because a histogram is a count over a range.

While a transfer is set (levels, a curve, or any stretch other than Original) a second, fainter trace is drawn over the first. It shows where that transfer sends the counts: how much of the output range the setting is using, and whether it is clipping. The input distribution alone cannot say either. Because a bin is an interval, each one's counts are spread across the output interval it lands on. The ghost is therefore smooth, where a naive mapping would comb into spikes at the values the transfer happens to hit.

Original draws no ghost. It is the state a scene opens in, and the solid trace already stands for it. It is also not the identity. Original maps the storage type's whole range. A 16-bit file holding Landsat values between about 7,000 and 30,000 uses barely a tenth of the display under it, while a neutral Levels or any authored stretch uses nearly all of it. That is where the large brightness jump on switching away from Original comes from. The ghost matters because, on a normalised value axis, two linear transfers differing only in scale draw exactly the same shape, however different the picture looks.

Black point, midtone and white point per channel, and a freehand transfer curve under the same histogram: click the curve to add a point, drag to shape it, double-click a point to remove it. From Equalise seeds the curve with the shape histogram equalisation would produce, then freezes it. Equalisation recomputes as the view moves. A curve does not, and it can be carried to another scene. Straighten returns it to the identity.

Offered when the layer is recognised as an elevation model. An illumination dial (the handle is the sun's position on the sky dome seen from above, centre overhead, rim at the horizon) with vertical exaggeration, multi-directional light, cast shadows, skylight, and a hypsometric, slope or aspect tint over any colormap. A live thumbnail follows every control. The full-resolution pass runs only on Apply.

Compiles the same expression language as Band Math to a Metal shader at runtime and draws the raster you are on as it, false-coloured through a colormap with a legend. Display only, no file. Use it when what you want is a look, and Band Math when you want a layer.

A reflectance transformation image holds, per pixel, a function of the light direction in place of a colour. A rock art panel or an inscription photographed under a dome of lights can be lit from any side afterwards. Open a polynomial texture map (.ptm, in the LRGB or the RGB layout) and it converts on the way in and opens relit. Open a .lp light file and the window of its first photograph offers to fit one from the whole capture, in Malzbender's biquadratic polynomial (six terms, one colour or per channel) or in hemispherical harmonics to order two or three (nine or sixteen terms per channel). The fit runs in the Jobs panel. Its result opens as its own window, with the surface normal from photometric stereo beside the coefficients.

The Relight section sits first in the Bands inspector for such a raster: the same dial as the relief shading, azimuth and elevation, and the mode. Relight evaluates the function under the light. Specular enhancement adds Malzbender's highlight from the stored normal. Diffuse gain scales the polynomial's curvature about its own maximum to deepen relief. It is a polynomial-only mode. Normals draws the surface normal as a colour. A light change is a uniform write and a frame. The light and the mode are saved in the project.

One control, five ways of making a polygon: Click Corners (double-click or click near the first corner to close), Free Draw (hold and trace, closing when you let go), Rectangle, Ellipse, and Select Similar Pixels, which lets the image decide where the edge is. The tick in the menu marks the live mode. Unticking it puts the tool away. Each polygon becomes its own layer, carries its ground area, and can be renamed, given a colour, edited vertex by vertex, or saved as GeoJSON. Marking one Use as Processing Region limits jobs to it. Region polygons draw in graticule cyan.

Press a pixel and drag away from it. The circle that opens under the pointer sets how much ground the selection may cover. It also sets how much of the range around that seed pixel counts as similar. Both widen the further you drag. What grows inside it, four-connected back to the seed, is drawn as a dotted outline while the button is held. Letting go commits that outline as one ordinary polygon, editable, nameable and usable as a processing region like any other.

Match pixels by has two settings, and they are different measurements. Value is how far the numbers are apart, scaled by each band's own spread. It suits one band or a stack of one quantity: a DEM, a conductivity grid, an index. Spectral shape compares the shape of the spectrum and ignores brightness. Sunlit and shaded ground of one material stay together, and it is the right choice for imagery.

Keeps the image inside or outside a polygon and makes the rest nodata. The inside case crops to the polygon's bounding box. The result is the shape and its surroundings, with no black expanse.

Materialises the provider's own calibration equations, the ones the delivery publishes, turning digital numbers into radiance, reflectance or brightness temperature. Scale and offset in both conventions, solar-elevation reflectance, Planck brightness temperature, Sentinel quantification, Maxar gain and bandwidth, DIMAP divisor gain, and lookup tables. Per-band quantity, units, equation and coefficient provenance are written into the result and shown in Info.

Estimates the atmospheric path radiance from the darkest pixels in each band and subtracts it. The scene percentile, the bounded estimation sample and whether negatives are clamped are all yours to set. The output identifies the estimated dark value and states explicitly that no mask was applied.

Estimates haze per location from a locally searched dark-object field (Makarau, Richter, Müller and Reinartz 2014), then subtracts it from each band scaled by how much that band sees haze. Dark-object subtraction takes one value per band for the whole scene. This takes a field, and haze that thickens across a scene is removed where it is. The coarse fields are folded in one streamed walk and upsampled per row inside the transform. No full-resolution haze field is ever held, and a scene of any size costs one row block.

Subtracts thin cirrus using the sensor's own 1.38 µm channel (Gao and Li 2017). At that wavelength the water-vapour column absorbs all surface signal, and the band records scattering alone. One relation is fitted per band for the whole scene. Every band of the source is carried through whether corrected or not. The output is still the scene.

Writes the tone curves you are looking through into a new raster. Every display channel is put through its own curve (stretch, levels, curve, histogram equalisation, decibel, whatever is in force). The result is written as a georeferenced GeoTIFF at full bit depth, either in the source's own data type over that type's whole range or as Float32 from 0 to 1. The 4096-entry table the shader samples is coarser than the data. The curve is evaluated per sample instead, and the file comes out finer than the screen was.

Removes the brightness difference between slopes facing the sun and slopes facing away. Reflectance can then be compared across terrain. SCS+C, C-correction, Minnaert with the constant fitted per band, and the statistical-empirical rotation. Sun zenith and azimuth are auto-filled from acquisition metadata where the product states them, and are editable otherwise. Slope and aspect come from a Horn 3 × 3 operator.

Resamples the raster onto a different coordinate reference system, writing a new file on the target grid. Every pixel is interpolated, and the output is a new set of measurements. The result opens as a new document.

Warps this raster onto another loaded raster's exact grid: same CRS, same extent, same pixel size. The two can then be compared or differenced cell for cell. Nearest, bilinear, cubic, cubic-spline, Lanczos, average or mode, chosen explicitly.

Takes a spectral selection, a rectangular window, or both, and writes them out without interpolating anything. Bands keep their source order and their values are the file's own.

Fits a transform from points whose position you know in both images (affine, second- or third-order polynomial, or thin-plate spline) and warps the raster through it. The subject is the active layer. The reference is whatever is already correctly placed beneath it. Each point is one feature clicked twice.

Aligns this raster to a reference by measuring the shift between them with GPU phase correlation, then warping once onto the reference's grid. Reports RMSE and peak ratio before and after.

Removes the terrain displacement in an image taken at an angle, using the sensor model the delivery carries (RPC or ground control points) together with a DEM. Target CRS, pixel size, interpolation and maximum transformer error are all explicit.

Injects the detail of a higher-resolution guide band into coarser bands. A multispectral scene takes the sharpness of its panchromatic one and keeps its colour relationships. Methods are SFIM, Brovey, GSA and MTF-GLP-HPM. A standard step before segmentation, object-based classification and any work where the boundary matters more than the pixel.

Guide data decides what the detail is taken from. Band n is the ordinary case for a panchromatic guide. Mean of all bands is the cheap fallback. Fitted per band (least squares) solves least squares from the guide's bands onto each coarse band, then applies those weights at full resolution. Each guide band is first degraded the way the chosen method degrades its guide. The third option makes fusion possible where no single guide band resembles the band being sharpened. MODIS bands 1 and 2 at 250 m against the 500 m shortwave bands are such a case. Snow is bright in red and dark at 1640 nm, and a single red guide injects every boundary upside down. That is worse than not fusing at all. The fit is free to weight a guide band negatively, and that case needs exactly this. It is offered for SFIM and MTF-GLP-HPM only, and only when the guide has more than one band. GSA and Brovey already fit or normalise a guide of their own, and a second fit underneath has no meaning.

Removes the periodic banding a pushbroom sensor's detector array leaves, by notching that frequency out of the Fourier transform or by subtracting a row or column profile. Stripe direction, centre frequency and bandwidth are all stated.

Smooths the image by suppressing high spatial frequencies, either through a GPU FFT or by an Accelerate Gaussian convolution.

Keeps the fine detail and removes the broad brightness variation, either as the bare residual or added back as sharpening with an amount you enter.

Enhances local contrast by unsharp masking, by guided filtering, or by Richardson–Lucy deconvolution against a Gaussian point spread function. Guided filtering preserves edges where unsharp masking haloes them.

Joins several rasters into one, choosing the output grid, where the seam runs between overlapping scenes, and how they are blended across it. Two seam solvers: least-cost path routed through pinned waypoints, and graph cut, which can isolate an island. Four join modes: hard, feathered with a linear or cosine ramp, distance-weighted, and seamless multi-band Laplacian. A live preview composites on a decimated grid and the seam can be edited by hand. Scenes in different coordinate systems are warped once into a target CRS you choose.

Brings one raster onto another's radiometric scale, by histogram specification, linear optimal transport matching mean and full covariance, global regression, LIRRN, or IR-MAD. Two dates or two sensors can then be compared.

Combines this raster's bands into one new layer, as a named index resolved from its spectral roles or as an expression you write. Presets cover NDVI, NDWI, NDMI, NBR, NDSI, NDRE and tasseled-cap brightness, greenness and wetness, resolved against whatever bands the image carries. A custom expression takes an arbitrary formula in b1…bN with arithmetic, comparison, logical and bitwise operators, conditionals, clamp, min and max, logarithms and QA bit extraction. Expressions evaluate over the raw samples, and the result is a measurement.

Mean, standard deviation, minimum, maximum and optional percentiles per band, over the whole image or a chosen region. For 8- and 16-bit storage the percentiles are exact and come from the same pass as the moments.

The same statistics, per polygon, in a single pass over the raster however many zones there are. A pixel belongs to a zone when its centre falls inside.

Principal components, or the minimum noise fraction, which is PCA after whitening against an estimate of the noise. Both write one band per component.

The ASTER AST06 colour enhancement. Three bands are rotated through the eigenvectors of their correlation matrix, each component is stretched to the same spread, and the rotation is undone. The colour axes stay where they were, and the correlation between them is gone. Writes red, green and blue at the target mean and standard deviation (127.5 and 50 by default, the ATBD's), uncorrelated. It is a visual aid, and the values have no physical units. The sheet previews the stretch on the image as its controls move, at every zoom, and writes nothing until Run. For a photograph the stretch can work in YUV or L*a*b* in place of RGB, as DStretch's named enhancements do. An emphasis wheel puts more of the contrast on one hue axis (reds and cyans, or yellows and blues) or on lightness.

Both commands work from a named set of reference spectra. There are two ways to get one. Derive a set from the selected polygons takes the mean or median of the pixels inside each polygon you have drawn over ground you know to be a single material, names each signature and its material class, and optionally adds a photometric shade endmember. Import Library… reads an external spectral library and convolves each spectrum onto this sensor's own bands through a Gaussian response built from each band's stated width. A band that states no width falls back to interpolation, and the set records which was used band by band. Whether the grid came from the sensor's published centres or from the wavelengths this particular file happens to carry is written into the set's provenance, because those are different claims. Sets can be deleted from the sheet that made them.

Solves each pixel as a mixture of the set's materials and writes the fractions. That is how a 10 m or 500 m pixel says how much bare soil, green vegetation or snow it holds. There are three models. The choice between them is about endmember variability: one signature per material rarely holds over a whole scene.

Single model puts every signature in one matrix. It is right when the set was derived from this scene, because then the compromise was made deliberately over ground you chose. MESMA solves every pixel against every candidate model and keeps the winner: fewest materials first, then lowest residual. Materials per model bounds the complexity, and Largest residual bounds admissibility. AutoMCU solves each pixel many times, drawing a different signature from each material's bundle. It reports the mean and the standard deviation, and a fraction arrives with an error bar. Fractions sum to one is the constraint that makes a shade endmember identifiable.

Labels each pixel with the reference spectrum it points most nearly at. It measures the angle between the pixel's spectrum and each reference and ignores their lengths. Illumination, slope and shade do not move the answer the way a distance measure would. The angle is written beside the class, and a confident match and a distant one are distinguishable.

Only match within an angle leaves a pixel unclassified beyond the largest angle you set. The angle is still written, and “nothing was within 5 degrees and the nearest was 23” stays readable. Unbounded matching is right only where the set is known to span the scene. Remove the continuum first divides both pixel and references by their own convex hull and compares absorption-feature shape. That is usually better for minerals and worse for vegetation and land cover. It is a different measurement and is recorded as one. Write every reference's angle adds one band per reference. It is the only way to see that a pixel was nearly a tie.

Compares two dates and reports both the classes and the continuous surface behind them: differencing, log-ratio for radar, index difference including dNBR burn severity, or IR-MAD's no-change probability as the threshold.

Differences two epochs of terrain and reports the volume of soil lost and gained, with the uncertainty that makes the figure defensible: a level of detection measured from ground you mark as stable, a probability of real change per cell, and a spatial-coherence pass that keeps a coherent patch of marginal cells and drops isolated spikes.

Turns a geocoded unwrapped interferogram into displacement, in metres, with the coherence that says which of it is a measurement. It has millimetre precision at 30–80 m posting and resolves slow motion: creep, subsidence, a slope moving before it fails.

Writes the steepness of an elevation model as a raster, in degrees, percent or rise over run, optionally with the downslope bearing beside it. It uses Horn's 3 × 3 kernel, the same one the relief shading and the topographic correction use. A slope class and a hillshade cannot disagree about the same cell. Flat ground gets no aspect.

Turns one band of continuous measurements into named classes by a table of breaks, writing a class raster with an attribute table. Where the breaks fall is the whole of what the output says, and a threshold typed with nothing behind it is chosen blind. The band's own distribution is drawn behind the breaks.

The New Zealand Land Use Capability system. Intersects up to five factor layers (rock type, soil, slope, erosion, vegetation) into inventory units, sieves them to a minimum mappable area, and looks each combination up in an editable rule table. The output is a unit map with a full attribute table: every unit's factor codes, its area, its LUC class 1 to 8 with its e/w/s/c subclass, and where each of its five statements came from.

The four jobs below are offered where Slope is, on a layer recognised as an elevation model. They share two rules. A radius is typed in metres and converted on this model's own cell size (10 m is 10 cells at 1 m and 20 at 0.5 m). The file records both. The window also has to be complete. A cell whose radius runs off the raster or into nodata is withheld. Every product carries a ring of one radius round the border and round every hole. Filling the ring would measure against surroundings that are partly imagined, and these are measurements. RVT mirrors its model outward at the edge. Here the margin is withheld, and the help says so.

Writes a cell's height against its surroundings: the local relief model (the height minus the mean within a radius, Hesse 2010, the same number as a topographic position index at a filled window), the deviation from mean elevation (that difference over the window's standard deviation), the multi-scale topographic position (the deviation of largest magnitude across three scale ranges, as three bands) and the multi-scale relief model (Orengo and Petrie 2018). Archaeologists run these on airborne LiDAR to see banks, ditches and mounds under the general lie of the land. At the same radius in millimetres, they find a petroglyph in a photogrammetry model of a rock panel.

Writes the second derivatives of an elevation model on the same 3 × 3 grid as Slope: profile curvature along the slope, plan curvature across it and the Laplacian, all in 1/m (Zevenbergen and Thorne 1987), or Riley's terrain ruggedness index, the root of the summed squared height differences to the eight neighbours, in metres. Curvature is how a bank, a ditch and a lynchet differ from the hillside they sit on. Ruggedness is the local roughness a ploughed-out barrow and a stone setting both raise.

Writes what a cell can see of the sky and what it looks down on, from the ray march the relief shading already runs: positive openness (90° minus the mean horizon angle within a radius, the horizon allowed below level so a summit exceeds 90°) and negative openness (the same on the inverted model, so a pit does); Zakšek's sky-view factor and its anisotropic form (the sky a cell sees, 0 to 1, optionally weighted towards a main direction); Hesse's local dominance (the mean angle at which an observer 1.7 m up looks down on the ground between two radii); the openness difference (positive minus negative, marched both ways and written as one signed product, convex ground above zero and hollows below); Kennelly and Stewart's sky illumination (the cell's own tilted surface lit by a uniform or a CIE overcast sky, each direction's sky above its horizon weighted by how squarely the surface faces it, 1 on open flat ground); and the sunlit mask, 1 where the sun clears the horizon in its direction. Openness and the sky-view factor are the layers the Visualization for Archaeological Topography is built from. The difference is the relief layer of the colour relief image map.

Writes a hillshade from every direction at once, one band per azimuth evenly spaced from north (Hobbs 1999). It uses the same Horn gradient as Slope and the relief shading, clamped at zero because a hillshade is light received. A single beam hides whatever lies parallel to it and drowns whatever faces it. Sixteen miss nothing. The first three principal components of the stack (Devereux, Amable and Crow 2008) carry almost all of it in one picture, and the panel can fit and write them beside the stack.

Assembles an archaeological composite as a stack of ordinary layers from the products above, running the jobs it needs through the Jobs panel and grouping the results under the recipe's name. Eight recipes, transcribed from RVT's own sources. VAT (general terrain): a hillshade from 315° at the bottom, the slope in luminosity over it on an inverted grey ramp so steep ground is dark, positive openness in overlay and the sky-view factor multiplied on top, each cut at the published range. VAT (flat terrain) lowers the sun to 15°, narrows every cut, widens the radius to 20 cells and skips the nearest 40 % of it so noise does not read as relief. VAT (combined) lays the general stack, as one picture, at half over the flat one. Colour relief image map puts slope on an orange–red ramp with the openness difference laid over it in luminosity and then overlay. Convex ground lightens and hollows darken. Prismatic openness lays positive and negative openness at half each over the slope, with no light direction. e4MSTP overlays RVT's MSTP picture on slope in red, with two composited pictures multiplied between them. The last two are pictures of the multi-directional hillshade: three directions as red, green and blue, and the PCA of sixteen as RGB, each component cut between its 2nd and 98th percentiles when it is made.

Finds groupings in the data without any training: K-Means with k-means++/maximin seeding, ISODATA with adaptive split and merge, or a Gaussian/Mahalanobis mixture. All deterministic and multicore, with an optional Metal backend for the assignment step. Images too large for memory are fitted on a decimated sample and assigned block by block. The output is a class raster with editable per-class colours and per-class ground areas in the Classes inspector.

A class is a name, a colour and a set of drawn polygons. Draw over ground you know, select the polygons and press Assign, or right-click a polygon to file it directly. Each row carries its polygon count and its pixel count. The pixel count is the figure that matters. A class needs more pixels than bands to fit a covariance at all, and ten to a hundred times that for the estimate to be stable. The row turns amber when it is thin for the bands you have selected. The advice names both remedies: draw more, or select fewer bands.

Every band is used unless you say otherwise. Untick the ones that carry no information for these classes. A thin class over sixty bands is ample over six, and band selection is often the cheaper of the two fixes. Once a separability check has run, Rank orders the bands by how well each one separates the weakest pair, scored one band at a time. Because bands are correlated, those scores do not add up to the matrix. Read them as a ranking.

A run that honoured no mask is a fact about the result. The Quality mask section states whether one was found beside the layer, whether or not there is one. With Reject masked training pixels on, cloud, shadow and fill are dropped from the training pixels only. The map still covers them. Masking the output is a separate decision, and it changes every area figure.

Scores every pair of classes before anything is fitted, as Jeffries-Matusita or transformed divergence, and reports each pair as separable, marginal or not distinct. Two classes that overlap in feature space cannot be told apart by any classifier, however good. This check says whether the problem is the method or the classes. The matrix exports as CSV, one row per pair, and the list filters to the problem pairs.

Assigns a class to every pixel from the training polygons you have labelled. Random Forest is the workhorse. It is an ensemble of trees, each on a bootstrap sample with a random subset of bands per split. It makes no assumption about the shape of a class and tolerates correlated bands and some mislabelled pixels. Maximum Likelihood fits a full covariance ellipsoid per class and is still the best where classes really are roughly Gaussian and there are pixels enough to estimate one. Minimum Distance needs very few training pixels. Spectral Angle Mapper ignores brightness. The forest exposes trees, maximum depth and a fixed seed. The same training set gives the same map every run. Change the seed only to see how much the answer depends on the bootstrap.

Blocked cross-validation, under Validation, refits the classifier with whole blocks of ground held out. A held-out pixel's neighbours were then not trained on. A random split over training polygons reports an accuracy inflated by up to 28 % by spatial autocorrelation. The blocked figure corrects it. Derive the block size from the data fits a variogram per band over the training pixels and takes the median range. It reports the size it used as a ground distance, and refuses where the pixels are too few or too clustered to fit one. Treat the derived figure as a floor. A block edge set by hand must be larger than your polygons. Otherwise a polygon straddling a boundary puts near-duplicate pixels on both sides, and the optimism returns.

The assessment of the map itself, on an independent sample. Choose how many sample units and how they are allocated (proportional to area, equal per class, or a floor on rare classes with the rest spread by area) and generate the sample. The points are drawn by reservoir sampling in one pass. This costs the same on a 500-megapixel map as on a small one. Then label each point from what you know of the ground, stepping through them, and read the assessment: overall accuracy, quantity disagreement, allocation disagreement and kappa, with the confusion matrix and a summary that both export.

Reads a towed electromagnetic-induction logger file and produces three georeferenced products in one window: a point raster of what was measured, an interpolated surface per channel, and optionally an inverted conductivity-against-depth stack. The panel lays the pipeline out as a stack of stages. Every filter is a row that switches off, and the pipeline re-runs on any change. Excluded records stay visible in grey, and you can see where a filter bites. Stages run in order: parse and project, acquisition geometry, motion filter, range filter, ferrous interference, along-track despike, rolling mean, GNSS latency, sensor offset, instrument drift, temperature standardisation, and calibration against reference profiles.

Takes a logger export and produces a ranked soil-pit plan. In one act it grids the survey, inverts it, groups the ground into management zones and sites the pits. Calibration wants ordinary ground, training wants unbiased coverage, and a diagnostic pit wants the anomalies. The three purposes are kept apart, and a site never carries two.

Groups the inverted survey into substrate units and draws them as polygons. Two places can have identical shallow readings and completely different profiles, and it is the profile that decides rooting depth and drainage. For that reason it clusters the whole column: log conductivity per layer, plus where the profile steps and by how much.

Samples the posterior distribution of the layered model under one sounding. The width of each layer's interval says how well these coils constrain it. Click a cell on the map to choose the sounding and set how many layers to ask the readings to support. Past a factor of ten, the data do not resolve that layer at all, however confident a single-model inversion looks.

Turns substrate units into a Land Resource Inventory factor layer. Each unit is listed with the evidence it was built from (the depth of its strongest conductivity step, its layer conductivities, its area and how confidently it was assigned) and you give each one a class from the handbook's own coding.

A ray-marched view of a layered conductivity model, with an orbit camera, a colour map and an opacity curve for picking out a range of values, optional depth-slice shading, the survey track, and depth posts scaled to the survey's own coils. Two guillotine blades are objects in the view. Drag an end post to cut, or click inside the volume to bring both blades to that point. Reset parks them at the edge. A section beneath stays linked to what the blades expose.

Searches by area, date window and cloud cover, then downloads the chosen bands as a tracked job. Five sources: Earth Search on AWS, Microsoft Planetary Computer, CNES GEODES (the SPOT World Heritage archive), the Alaska Satellite Facility (Sentinel-1 and NISAR products, including the unwrapped interferograms InSAR Displacement reads) and LINZ. Landsat is free on Planetary Computer and requester-pays on Earth Search. GEODES needs a free API key for downloads and none for searching. MODIS is on Planetary Computer only, grouped under Reflectance, Vegetation, Temperature and fire, and Snow and ice. It arrives on the sinusoidal grid. Reproject is usually needed before combining it with a UTM or NZTM scene, to a local projected CRS where the tile straddles 180°.

Because LINZ publishes a static catalogue with no search service behind it, this source is selected by map sheet. The world map becomes the Topo50 1:50,000 sheet grid. Click the sheets you want, or drag a band across the grid to take every sheet it touches. Find Surveys then asks the bucket which aerial surveys hold tiles for them. The results are surveys, newest first. Over one sheet, that list is a time series of the same ground across decades. Choosing a survey draws its tiles. Click or drag on the map or the list, which are one selection. Hovering names the sheet the way LINZ printed it, and the tile by its filename.

A stalled or dropped transfer is retried a few times on its own and resumed where the archive allows it. If it still fails, the row states what went wrong in one sentence. It offers Retry where trying again can help, as after a timeout or a lost connection. A refusal such as HTTP 403 would answer the same way twice, and it is not offered one.

Writes a GeoTIFF of source measurements through GDAL, carrying any crop through as a source window and persisting attribute-table edits. The rendered picture exports separately as PNG, HEIC or OpenEXR through Image I/O. Every GeoTIFF the app writes is ZSTD-compressed and tiled. A file written without its georeference would look complete, and a writer that cannot copy the georeference throws.

A .hachureproj file records the arrangement: every project in the window and its coordinate system, the layer tree with its groups, blend modes and any levels a preset put on a group, display settings including a reversed ramp, the light and mode of a relit raster, polygons, annotations, training classes and the plans a survey produced. The images themselves are referenced. Keep them where they are. Opening a project rebuilds the window. Anything that has moved since is reported by name. Saving offers to copy any layer still living in the temporary folder into a folder beside the project.

A processing job writes its result into a temporary file and opens it as a layer. That makes the Processing menu feel immediate. It also means closing the window would throw the result away. Closing a window holding such layers lists them first, with a tick and a format for each, and Save, Don't Save or Cancel. The formats are all scientific containers: GeoTIFF, Cloud-Optimised GeoTIFF, KEA and ERDAS Imagine.

Add to Photos renders what is on screen and files it in a Hachure album, created the first time. It is the same picture the rendered export writes, display bands through their tone curves. The file says the curve is baked into its pixels, and the scene's centre goes in as GPS. A photo library has nowhere to put a float HDR image, and a GeoTIFF of measurements does not belong in one. Only PNG and HEIC are offered.

macOS gives an app exactly the files chosen in an open panel and nothing beside them. Several formats are not one file: a shapefile keeps its projection in a sibling .prj and its attributes in a .dbf, and a satellite delivery is a manifest beside its band files. There is no way to widen the permission without you choosing the wider thing. When a companion cannot be read, Hachure asks for the enclosing folder.

A 10–80 % ceiling on unified memory, 50 % by default, that every whole-image job checks before reading a pixel. An over-large job fails in a second with the required and available figures on screen. Run Anyway is offered, with Cancel as the default. Under memory pressure from elsewhere on the Mac, every window releases the caches it can rebuild. GDAL's block cache is held down until the pressure clears. Nothing on screen changes.

The Organiser keeps each scene's description and thumbnail between launches, under a byte cap you set. A folder of two hundred deliveries lists at once on the second opening. Clear empties it, and the next opening reads everything again.

Twenty sensors with per-band name, aliases, common name, centre wavelength, FWHM, native ground sample distance, spectral role, polarisation and thermal constants, plus versioned calibration profiles keyed on processing level, collection, baseline and acquisition date. It is user-editable: sensors can be added or overridden, and edits are validated before they are saved. The centres are the agencies' measured ones. ASTER's, for instance, are derived from JPL's published spectral response functions, and a test holds them to the nanometre.

An added raster must share the base's CRS. Reproject one onto the other's CRS, then add it again. If they share a CRS but sit slightly apart, that is a registration problem. Try Co-registration first.

Regression and IR-MAD normalisation, every change-detection method, and DEM of Difference all compare co-located pixels. The two rasters must have the same size, CRS and pixel size. Resample to Match Layer puts one on the other's exact grid. Use nearest for anything thematic.

Raw 12- and 16-bit radiometric data is near-flat without a stretch, and Hachure applies none on open. Percentile clip with a modest gamma is the usual first move. The stretch is display exposure only and changes nothing stored.

Check the spectral roles. A named index resolves against the bands the stack says are NIR and Red. If those are wrong, the arithmetic still succeeds and produces a convincing surface of the wrong two bands. Correct the role on the band's row, then re-run.

That is the complete-window rule. A cell whose radius runs off the raster or into nodata is withheld. Every windowed product carries a ring of one radius round the border and round every hole. The sky products march 100 cells by default and withhold most of a small tile. Run them on a sheet, or accept the ring. The alternative would be to measure against surroundings the model does not have.

White across the whole panel means the coefficients are on a different scale from the one the shader expects. A raster fitted by this app records its sample scale, and a PTM carries its own. White therefore points at a file made elsewhere. Black under diffuse gain comes from the mathematics. A gain of three under a 20° light reaches zero well before the horizon. Raise the light or lower the gain.

The .prj and .dbf are separate files and the sandbox grants only what you selected. Name a Workspace folder that contains them, or answer the folder prompt when it appears.

Hachure reads what its curated GDAL driver set reads. HEIC, HEIF and AVIF are left out of it on licence grounds. They are converted through the system's own image framework before anything else looks at the file, whether they arrive through File ▸ Import from Photos, from the Finder, at launch, or inside a folder. The converted copy is written into Hachure's own Application Support folder. Nothing is written beside your data. An alert says that it happened, why, and where the copy went.