Matlab Analyze Spectral Images

MathWorksOfficialCreated by MathWorks

Skill
Tool integration
Science and research
BSD-2-Clause
No issues found

From the author

Work with hyperspectral and multispectral images in MATLAB. Covers reading/writing (ENVI, NITF, TIFF, Sentinel-2, Landsat, ASTER), ECOSTRESS spectral libraries, processing (calibration, atmospheric correction, denoising, sharpening, dimensionality reduction, endmember extraction, unmixing, target/anomaly detection, spectral indices, segmentation), labeling (Spectral Image Labeler app, ground truth objects, automation algorithms), and deep learning (pixel classification CNNs, unmixing autoencoders, transfer learning). Use when reading, writing, processing, analyzing, classifying, labeling, or applying deep learning to hyperspectral or multispectral images.

— the description by MathWorks

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill provides a comprehensive interface for working with hyperspectral and multispectral data within the MATLAB environment. It covers the entire lifecycle of spectral imaging, including data ingestion from various formats (ENVI, NITF, TIFF, Sentinel-2, Landsat, ASTER), radiometric calibration, atmospheric correction, and advanced analytical techniques such as spectral unmixing, anomaly detection, and spectral index calculation. The skill also supports modern deep learning workflows, including pixel classification with 3D CNNs and semantic segmentation, as well as interactive labeling via the Spectral Image Labeler app.

What it does

Reads and writes industry-standard spectral image formats like ENVI, NITF, and GeoTIFF.
Performs radiometric calibration and atmospheric correction on raw spectral data.
Executes advanced spectral analysis including endmember extraction, unmixing, and target detection.
Computes standard and custom spectral indices like NDVI and EVI using optimized built-in functions.
Facilitates spectral-aware image segmentation and clustering.
Supports deep learning pipelines for pixel classification and semantic segmentation of spectral cubes.

When to use it

When processing hyperspectral or multispectral satellite or airborne imagery.
When performing quantitative analysis on spectral signatures using ECOSTRESS libraries.
When building deep learning models for spectral image classification or unmixing.
When needing to compute spectral indices or perform anomaly detection on hypercubes.
When managing large-scale spectral data with block processing requirements.

Advantages

Optimized built-in functions for spectral indices and processing, avoiding manual computation errors.
Seamless integration with MATLAB's deep learning and image processing toolboxes.
Native support for complex spectral data structures and metadata management.
Includes specialized apps for interactive labeling and visualization of high-dimensional data.

Usage examples

“Read a hyperspectral image from an ENVI file and calculate the NDVI.”: The agent provides code to load the file using imhypercube and compute the index using the built-in ndvi function.
“How do I perform anomaly detection on my hyperspectral cube?”: The agent provides the syntax for anomalyRX to identify anomalous pixels within the spectral data.
“Extract endmembers from my hyperspectral data.”: The agent provides a workflow using countEndmembersHFC to determine the number of endmembers followed by nfindr for extraction.
“How can I visualize my multispectral cube as a true-color image?”: The agent provides the colorize function syntax with the appropriate method parameter for multicube visualization.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/matlab/matlab-analyze-spectral-images/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip matlab-analyze-spectral-images.zip -d .agents/skills

Documentation by MathWorks

Spectral Imaging Skill

Generate MATLAB code for working with hyperspectral and multispectral data using the Hyperspectral Imaging Library for Image Processing Toolbox. Requires desktop MATLAB (not MATLAB Online or MATLAB Mobile) and the Hyperspectral Imaging Library add-on.

When to Use

Reading/writing hyperspectral or multispectral images (ENVI, NITF, GeoTIFF, TIFF, ASTER, Sentinel-2, Landsat)
Reading ECOSTRESS spectral library signatures (readEcostressSig)
Creating hypercube/multicube objects, band selection, cropping, block processing
Radiometric calibration, atmospheric correction, spectral correction
Denoising, sharpening, dimensionality reduction (PCA, MNF)
Endmember extraction, spectral unmixing, abundance estimation
Spectral matching, target detection, anomaly detection
Computing spectral indices (NDVI, EVI, custom)
Segmentation (spectral clustering, superpixels, ISODATA)
Labeling with the Spectral Image Labeler app
Creating ground truth objects for spectral labeling
Deep learning classification or unmixing of spectral images
Semantic segmentation of multispectral satellite imagery (DeepLabV3+ with multi-channel input)

When NOT to Use

General image reading/writing (use imread/imwrite)
Non-spectral 3-D volume data (use Medical Imaging or Image Processing Toolbox)
General image classification not involving spectral cubes (use Deep Learning Toolbox)
Non-spectral semantic segmentation of RGB/grayscale images (use Computer Vision Toolbox)
Point spectra or spectral libraries without spatial dimensions

Prerequisite Check

If you suspect the Hyperspectral Imaging Library might not be installed (e.g., user reports errors), verify with exist('imhypercube','file'). If not installed, suggest installing via Add-On Explorer and fall back to multibandread/imread. Do NOT run this check if you are confident the library is available.

Quick Start

% Read hyperspectral data
hcube = imhypercube("scene.hdr");

% Visualize — colorize syntax (NO name-value for basic usage):
falseColor = colorize(hcube);              % false-colored image (default, 3 most informative bands)
rgb = colorize(hcube, Method="rgb");       % true RGB (hypercube ONLY — needs R/G/B wavelengths)
cir = colorize(hcube, Method="cir");       % color infrared (hypercube ONLY)
custom = colorize(hcube, [30 20 10]);      % false color from specific band INDICES (not wavelengths)
imshow(falseColor)

% Radiometric pipeline
hcube = dn2radiance(hcube);
hcube = radiance2Reflectance(hcube);
hcube = fastInScene(hcube);

% Compute NDVI
ndviMap = ndvi(hcube);

% Endmember extraction and unmixing
numEM = countEndmembersHFC(hcube);
endmembers = nfindr(hcube, numEM);
abundance = estimateAbundanceLS(hcube, endmembers, Method="fcls");

% Anomaly detection
rxScore = anomalyRX(hcube);

% Segmentation
L = hyperseganchor(hcube, 8);

% Launch viewer app
hyperspectralViewer(hcube)

% Launch labeler app (Since R2026a)
spectralImageLabeler(hcube)

Essential Syntax (use without reading reference files)

colorize — Visualization

% HYPERCUBE (supports Method, 2 outputs, ContrastStretching default=false)
falseColor = colorize(hcube)                        % false-colored (default)
rgb = colorize(hcube, Method="rgb")                 % true color RGB
cir = colorize(hcube, Method="cir")                 % color infrared
custom = colorize(hcube, [b1 b2 b3])               % 3 band indices
[img, indices] = colorize(hcube)                    % also get band indices used
img = colorize(hcube, ContrastStretching=true)      % with CLAHE

% MULTICUBE (NO Method param, 1 output, ContrastStretching default=true)
rgb = colorize(mcube)                               % true RGB (auto-selects by wavelength)
custom = colorize(mcube, [b1 b2 b3])               % false color from band indices
img = colorize(mcube, ContrastStretching=false)     % disable contrast stretching

I/O and Band Management

hcube = imhypercube("file.hdr");                    % read ENVI/NITF/TIFF
mcube = immulticube("MTL.txt");                     % read Landsat (pass *_MTL.txt file)
mcube = immulticube("manifest.safe");              % read Sentinel-2 (pass manifest.safe file, NOT .SAFE folder)
mcube = immulticube("AST_L1T.hdf");                % read ASTER HDF
cropped = cropData(hcube, 30:110, 30:110);          % row indices, col indices (NOT [x y w h])
vnir = selectBands(hcube, Wavelength=[400 1000]);   % by wavelength range
subset = selectBands(hcube, BandNumber=[1 10 20]);  % by band index
cleaned = removeBands(hcube, Wavelength=[1350 1450]); % remove wavelength range
data = gather(hcube);                               % get M-by-N-by-C numeric array

% ACCESSING DIMENSIONS — NO NumRows/NumColumns/NumBands properties exist:
height = hcube.Metadata.Height;                     % number of rows (M)
width  = hcube.Metadata.Width;                      % number of columns (N)
bands  = hcube.Metadata.Bands;                      % number of spectral bands (C)
% For multicube: use mcube.BandSize(1,1) for height, mcube.BandSize(1,2) for width

% RESAMPLING vs SELECTING (multicube only):
mcubeUniform = resampleBands(mcube, 30);                       % RESAMPLE all bands TO 30m (keeps ALL bands, changes spatial size)
mcubeUniform = resampleBands(mcube, 30, Method="bilinear");    % interpolation: "nearest"(default),"bilinear","cubic"
mcube30m = selectBands(mcube, "DataResolution", 30);           % KEEP only bands already at 30m (DISCARDS others)
% When asked to "resample" or "match resolution" → use resampleBands
% When asked to "select" or "filter" bands by resolution → use selectBands

Spectral Indices (NEVER compute manually — use built-in functions)

% NDVI — use built-in ndvi() function (supports block processing natively)
ndviMap = ndvi(spcube);                             % direct numeric M-by-N output, range [-1,1]
ndviMap = ndvi(spcube, BlockSize=[512 512]);        % with block processing for large data

% Standard indices — use spectralIndices() (returns struct with .IndexName and .IndexImage)
indices = spectralIndices(spcube, "NDVI");          % single index → struct
indices = spectralIndices(spcube, ["NDVI","EVI"]);  % multiple → struct array
indices = spectralIndices(spcube, "all");           % all 16 supported indices
indices = spectralIndices(spcube, "NDVI", BlockSize=[512 512]);  % with block processing
% Supported: "NDVI","EVI","NDBI","NBR","MNDWI","MSI","GVI","OSAVI","SR","CMR","MCARI","MTVI","PRI","NDNI","NDMI","CAI"

% Custom formula — use customSpectralIndex() (3 required args: cube, wavelengths, @func)
indexImage = customSpectralIndex(spcube, [800 670], @(nir,red) (nir-red)./(nir+red));
indexImage = customSpectralIndex(spcube, [800 670 475], @(nir,red,blue) 2.5*(nir-red)./(nir+6*red-7.5*blue+1));
indexImage = customSpectralIndex(spcube, [560 1650], @(g,swir) (g-swir)./(g+swir), BlockSize=[512 512]);
% DO NOT manually extract bands and compute — always use these built-in functions

Detection, Matching, and Processing

% Target detection — method is POSITIONAL (3rd arg, not Name=Value)
target = resampleSignature(libData(1), hcube.Wavelength); % resample library spectrum to cube wavelengths
score = detectTarget(hcube, target, "ACE");         % "CEM","ACE","SignedACE","MF","GLRT","AMSD","OSP"
rxScore = anomalyRX(hcube);                         % higher = more anomalous

% Similarity scores — cube first, reference second (lower = better match)
score = sam(hcube, refSpectrum);                    % spectral angle mapper [0, pi]
score = sid(hcube, refSpectrum);                    % spectral information divergence

% spectralMatch — UNUSUAL ARG ORDER: library struct FIRST, then cube/spectrum
% DO NOT pass numeric vectors — first arg MUST be a struct with .Reflectance and .Wavelength fields
libData = readEcostressSig("vegetation.txt");       % returns 1-by-K struct array
score = spectralMatch(libData, hcube);              % returns M-by-N-by-K (one score map per signature)
score = spectralMatch(libData, hcube, Method="sam");% "sam"(default),"sid","sidsam","jmsam","ns3"
score = spectralMatch(libData, reflectance, wavelength); % compare against single spectrum (returns K-vector)
% Lower score = stronger match. Returns NaN when bandwidth overlap < MinBandWidth (default 300nm)
% For direct numeric comparison without library: use sam(hcube, target) or sid(hcube, target)

% Spectral clustering / segmentation (use INSTEAD of kmeans for spectral data)
L = hyperseganchor(hcube, numClusters);             % spectral-aware clustering (1 output)
[L, n] = hyperslic(hcube, K);                       % spectral superpixels (2 outputs)

% Block processing — callback receives OBJECT, must use gather()
result = apply(hcube, @(block) mean(gather(block), 3));

Workflow

For functions fully documented in the Essential Syntax and Common Mistakes sections above, generate code directly — those sections provide complete API signatures and constraints.

Read a reference file ONLY when:

Using a function NOT covered in Essential Syntax (e.g., sharc, empiricalLine, sharpencnmf, spectralImageLabeler)
Unsure about argument order or constraints for a complex workflow
The task involves the Spectral Image Labeler app or ground truth objects

Reference Files (consult only when needed):

references/spectral-io.md — I/O, band selection, cropping, block processing, colorize, enviwrite
references/spectral-calibration.md — Radiometric calibration, atmospheric correction, spectral correction, denoising, sharpening
references/spectral-analysis.md — PCA/MNF, unmixing, spectral matching, target/anomaly detection, indices, segmentation
references/spectral-deep-learning.md — Labeler app, groundTruthSpectralImage, deep learning pipelines
references/dl-classification.md — 3-D CNN, 1-D CNN, ground truth to training data workflows
references/dl-advanced.md — Unmixing autoencoder, transfer learning, DeepLabV3+ semantic segmentation

CRITICAL: Common Mistakes to Avoid

These are the most frequent errors. Violating any of these will cause runtime failures:

NEVER use hypercube() — deprecated. Use imhypercube() or geohypercube().
cropData(hcube, 30:110, 30:110) — row/col INDEX VECTORS, not [x y w h]
selectBands/removeBands require Name=Value: selectBands(hcube, BandNumber=[1 10 20]), removeBands(hcube, Wavelength=[1350 1450]). removeBands errors if range outside cube's actual wavelengths.
customSpectralIndex(spcube, [800 670], @(nir,red) (nir-red)./(nir+red)) — 3 args required
apply(hcube, @(block) mean(gather(block), 3)) — callback receives OBJECT, use gather()
inverseProjection(pcaCube, coeff) — data first, coefficients second
spectralMatch(libData, hcube) — struct from readEcostressSig FIRST. For numeric: use sam(hcube, target) directly.
L = hyperseganchor(hcube, 8) (1 output); [L, n] = hyperslic(hcube, 100) (2 outputs)
immulticube does NOT support ENVI. Sentinel-2: pass manifest.safe (NOT .SAFE folder). Landsat: *_MTL.txt. ASTER: .hdf.
spectralIndices(hcube, "NDVI") → struct with .IndexName/.IndexImage. Use ndvi(hcube) for numeric.
hyperslic with 3 bands: hyperslic(data, 100, IsInputDimReduced=true). Only params: NumIterations (max 30), IsInputDimReduced. NO Compactness.
detectTarget(hcube, target, "ACE") — method is POSITIONAL, not Name=Value
denoiseNGMeet: needs spatial dims > 16. Crop large images to ≤1024x1024 first.
flatField(hcube, [xmin ymin width height]) — rect vector, NOT index vectors like cropData
colorize(mcube) → true RGB (NO Method param!). colorize(mcube, [b1 b2 b3]) for false-color. colorize(hcube, Method="rgb") only works on hypercube.
resampleBands(mcube, 30) = resizes all bands to target res. selectBands(mcube, "DataResolution", 30) = keeps only bands already at that res.
NO NumRows/NumColumns/NumBands properties. Use hcube.Metadata.Height/Width/Bands. Multicube: mcube.BandSize(1,1)/mcube.BandSize(1,2).
NEVER manually compute indices. Use ndvi(spcube), spectralIndices(spcube, "NDVI"), or customSpectralIndex(spcube, [wl1 wl2], @(b1,b2) formula).
sam/sid refSpectra must be C-element VECTOR, not matrix.
assignData(hcube, rows, cols, bands, data) — 5 args with index ranges
enviwrite(hcube, "output") — hypercube ONLY, not multicube
groundTruthSpectralImage labelDefs: 6 columns required including AttributeList. Property: LabelDefinitions (plural).

Key Rules

Pipeline: DN → radiance → reflectance → atmospheric correction. Smile correction BEFORE denoising.
L2 products already have surface reflectance — DN pipeline is for L1 only.
Research datasets (Indian Pines, PaviaU) lack calibration metadata.
hypercube ONLY: denoiseNGMeet, hyperpca, hypermnf, inverseProjection, sharpencnmf, countEndmembersHFC, nfindr, ppi, estimateAbundanceLS, anomalyRX, sam, sid, spectralMatch, hyperslic, sharc, empiricalLine, reduceSmile
hypercube + multicube: dn2radiance, radiance2Reflectance, fastInScene, detectTarget, ndvi, spectralIndices, customSpectralIndex, hyperseganchor
+ numeric array: flatField, iarr, logResiduals, subtractDarkPixel
Deep learning: trainnet + dlnetwork (not trainNetwork). Use hyperpca before CNN.
Valid spectralIndices names: ndvi, osavi, sr, evi, gvi, ndni, pri, cai, mcari, mtvi, msi, cmr, nbr, ndbi, ndmi, mndwi (NOT "ndwi")
Labeler (R2026a+): groundTruthSpectralImage properties are read-only after creation

Files

ATTRIBUTION.mdAdded by AutoGineer1.1 KiB
LICENSELicence file added by AutoGineer1.4 KiB
SKILL.md14.5 KiB
manifest.yaml498 B
references/dl-advanced.md7.3 KiB
references/dl-classification.md7.7 KiB
references/spectral-analysis.md14.5 KiB
references/spectral-calibration.md10.0 KiB
references/spectral-deep-learning.md10.9 KiB
references/spectral-io.md17.0 KiB

references/dl-classification.md

Deep Learning Classification Workflows

Pixel classification using 3-D spatial-spectral CNN, 1-D spectral CNN, and ground truth to training data pipeline.

Workflow 1: Pixel Classification with 3-D Spectral-Spatial CNN (CSCNN)

%% Load and preprocess
hcube = imhypercube("indian_pines.dat");
gtLabel = load("indian_pines_gt.mat").indian_pines_gt;
numClasses = max(gtLabel(:));

% Dimensionality reduction via PCA
numPCBands = 30;
imageData = hyperpca(hcube, numPCBands);

% Normalize by per-band standard deviation
sd = std(imageData, [], 3);
imageData = imageData ./ sd;

%% Create spatial patches
windowSize = 25;
inputSize = [windowSize windowSize numPCBands];
[allPatches, allLabels] = createImagePatchesFromHypercube(imageData, gtLabel, windowSize);

% Filter labeled patches only
patchesLabeled = allPatches(allLabels > 0, :, :, :);
patchLabels = categorical(allLabels(allLabels > 0));

% Split: 30% train, 70% test
[trainIdx, ~, testIdx] = dividerand(size(patchesLabeled, 1), 0.3, 0, 0.7);
XTrain = permute(patchesLabeled(trainIdx, :, :, :), [2 3 4 1]);
YTrain = patchLabels(trainIdx);
XTest = permute(patchesLabeled(testIdx, :, :, :), [2 3 4 1]);
YTest = patchLabels(testIdx);

% Create datastores
dsTrain = augmentedImageDatastore(inputSize, XTrain, YTrain);
dsTest = augmentedImageDatastore(inputSize, XTest, YTest);

%% Define 3-D CNN (Custom Spectral CNN architecture)
layers = [
    image3dInputLayer(inputSize, Name="Input", Normalization="none")
    convolution3dLayer([3 3 7], 8, Name="conv3d_1")
    reluLayer(Name="relu_1")
    convolution3dLayer([3 3 5], 16, Name="conv3d_2")
    reluLayer(Name="relu_2")
    convolution3dLayer([3 3 3], 32, Name="conv3d_3")
    reluLayer(Name="relu_3")
    convolution3dLayer([3 3 1], 8, Name="conv3d_4")
    reluLayer(Name="relu_4")
    fullyConnectedLayer(256, Name="fc1")
    reluLayer(Name="relu_5")
    dropoutLayer(0.4, Name="drop_1")
    fullyConnectedLayer(128, Name="fc2")
    dropoutLayer(0.4, Name="drop_2")
    fullyConnectedLayer(numClasses, Name="fc3")
    softmaxLayer(Name="softmax")
];
net = dlnetwork(layers);

%% Train
options = trainingOptions("adam", ...
    InitialLearnRate=0.001, ...
    LearnRateSchedule="piecewise", ...
    LearnRateDropPeriod=30, ...
    LearnRateDropFactor=0.01, ...
    MaxEpochs=100, ...
    MiniBatchSize=256, ...
    GradientThresholdMethod="l2norm", ...
    GradientThreshold=0.01, ...
    ValidationData=dsTest, ...
    ValidationFrequency=100, ...
    Plots="training-progress");

net = trainnet(dsTrain, net, "crossentropy", options);

%% Predict and evaluate
dsAll = augmentedImageDatastore(inputSize, permute(allPatches, [2 3 4 1]), allLabels);
scores = minibatchpredict(net, dsAll);
prediction = scores2label(scores, categories(YTrain));

% Mask unlabeled pixels
prediction(allLabels == 0) = "<undefined>";
[M, N, ~] = size(imageData);
classMap = reshape(prediction, [N M])';

Workflow 2: 1-D Spectral CNN for Pixel Classification

%% Load data
hcube = imhypercube("scene.hdr");
dataCube = gather(hcube);
[M, N, C] = size(dataCube);

%% Load ground truth and extract labeled pixels
load("gTruth.mat", "gTruth");
labelMap = createLabelMapFromGTruth(gTruth, [M N]);  % user-defined helper

pixels = reshape(dataCube, [], C);
labels = reshape(labelMap, [], 1);
validIdx = ~isundefined(labels);
trainPixels = pixels(validIdx, :);
trainLabels = labels(validIdx);

% Reshape for 1-D CNN: C-by-1-by-1-by-numSamples
numSamples = sum(validIdx);
XTrain = reshape(trainPixels', [C 1 1 numSamples]);

%% Define 1-D spectral CNN
numClasses = numel(categories(trainLabels));
layers = [
    image3dInputLayer([C 1 1], Normalization="none")
    convolution3dLayer([7 1 1], 32, Padding="same")
    batchNormalizationLayer
    reluLayer
    convolution3dLayer([5 1 1], 64, Padding="same")
    batchNormalizationLayer
    reluLayer
    convolution3dLayer([3 1 1], 128, Padding="same")
    batchNormalizationLayer
    reluLayer
    globalAveragePooling3dLayer
    fullyConnectedLayer(numClasses)
    softmaxLayer
];
net = dlnetwork(layers);

%% Train
options = trainingOptions("adam", ...
    MaxEpochs=50, ...
    MiniBatchSize=256, ...
    InitialLearnRate=1e-3, ...
    Shuffle="every-epoch", ...
    Plots="training-progress");

net = trainnet(XTrain, trainLabels, net, "crossentropy", options);

%% Predict full image
XAll = reshape(pixels', [C 1 1 M*N]);
scores = minibatchpredict(net, XAll);
predLabels = scores2label(scores, categories(trainLabels));
classMap = reshape(predLabels, [M N]);

Workflow 3: Ground Truth to Training Data Pipeline

%% Load exported ground truth from Spectral Image Labeler
load("labelerSession.mat", "gTruth");

%% Load spectral data from the ground truth source
hcube = imhypercube(gTruth.DataSource.Filename);
dataCube = gather(hcube);
[M, N, C] = size(dataCube);

%% Convert polyshape ROIs to pixel label mask
labelNames = gTruth.LabelDefinitions.LabelName;
labelMask = zeros(M, N);

for j = 1:numel(labelNames)
    for row = 1:height(gTruth.LabelData)
        roiData = gTruth.LabelData.(labelNames(j)){row};
        if ~isempty(roiData)
            vertices = roiData.ROI.Vertices;
            mask = poly2mask(vertices(:,1), vertices(:,2), M, N);
            labelMask(mask) = j;
        end
    end
end

%% Create training arrays
pixels = reshape(dataCube, [], C);
pixelLabels = reshape(labelMask, [], 1);
validIdx = pixelLabels > 0;
XTrain = pixels(validIdx, :);
YTrain = categorical(pixelLabels(validIdx), 1:numel(labelNames), labelNames);

%% Or create patches for spatial-spectral CNN
patchSize = 25;
halfPatch = floor(patchSize/2);
[rows, cols] = find(labelMask > 0);

patches = zeros(patchSize, patchSize, C, numel(rows), 'single');
patchLabels = categorical.empty;
count = 0;

for i = 1:numel(rows)
    r = rows(i); c = cols(i);
    if r > halfPatch && r <= M-halfPatch && c > halfPatch && c <= N-halfPatch
        count = count + 1;
        patches(:,:,:,count) = dataCube(r-halfPatch:r+halfPatch, c-halfPatch:c+halfPatch, :);
        patchLabels(count,1) = categorical(labelMask(r,c), 1:numel(labelNames), labelNames);
    end
end
patches = patches(:,:,:,1:count);
patchLabels = patchLabels(1:count);

Common Patterns

Creating Datastores for Large Datasets

% Array datastores for patches
patchDS = arrayDatastore(patches, IterationDimension=4);
labelDS = arrayDatastore(patchLabels);
trainDS = combine(patchDS, labelDS);

% Use with trainnet
net = trainnet(trainDS, net, "crossentropy", options);

Evaluating Classification Results

% Predict and convert scores to labels
scores = minibatchpredict(net, dsTest);
YPred = scores2label(scores, categories(YTest));

% Confusion matrix
figure;
confusionchart(YTest, YPred);

% Overall and per-class accuracy
accuracy = mean(YPred == YTest);
cm = confusionmat(YTest, YPred);
perClassAcc = diag(cm) ./ sum(cm, 2);
fprintf("Overall Accuracy: %.2f%%\n", accuracy * 100);

Visualizing Classification Map

[M, N, ~] = size(imageData);
classMap = reshape(classIdx, [M N]);

figure;
imagesc(classMap);
colormap(jet(numClasses));
colorbar(TickLabels=categories(YTrain), Ticks=1:numClasses);
title("Classification Map");

Helper: Patch Extraction from Hypercube

function [patches, labels] = createImagePatchesFromHypercube(imageData, gtLabel, windowSize)
    [M, N, C] = size(imageData);
    halfWin = floor(windowSize / 2);
    padded = padarray(imageData, [halfWin halfWin 0], 0, "both");

    numPixels = M * N;
    patches = zeros(numPixels, windowSize, windowSize, C, 'single');
    labels = reshape(gtLabel, [], 1);

    idx = 0;
    for r = 1:M
        for c = 1:N
            idx = idx + 1;
            patches(idx, :, :, :) = padded(r:r+windowSize-1, c:c+windowSize-1, :);
        end
    end
end

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Security

No issues found by our static scan· No issues found

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

MathWorks

Publisher

MathWorks (Official)

Source

matlab/matlab-agentic-toolkit · skills-catalog/image-processing-and-computer-vision/matlab-analyze-spectral-images · commit 2a7f215

Licence

BSD 2-Clause License (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE.md)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by MathWorks or MathWorks.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download matlab-analyze-spectral-images.zip