fullFlowG.m

% Example full flow

% load images
    % animalTensor is an nr x nc x nScenario binary array,
    % where a 1 in spot (r,c,s) refers to a predicted presence in cell
    % (r,c) according to scenario s.
    load animalTensor
    % animOrig is an nr x nc array, where a 1 in spot (r,c) corresponds
    % to a present day presence in cell (r,c).
    load animOrig
    
% Compute pairwise similarities
    A = sum(animalTensor, 3) > 0 | animOrig;

    % create overlap with present day map and map pixel values for cosine
    novelPresence = 0.5;
    keptPresence = 1.0;
    novelAbsence = -1.0;
    keptAbsence = -0.5;
    
    animalValueTensor = double(animalTensor);
    animalValueTensor(animalTensor & animOrig & A) = keptPresence;
    animalValueTensor(animalTensor & (1 - animOrig) & A) = novelPresence;
    animalValueTensor((1 - animalTensor) & animOrig & A) = novelAbsence;
    animalValueTensor((1 - animalTensor) & ( 1 - animOrig) & A) = keptAbsence;
    pairMat = pairCosineSimG(animalValueTensor); % in codes/simMeasure
 
% Spectral Embedding

    kNearestNeighbors = 5;
    % Typically need 1 more eigenvector than embedding dimension.
    numEigenvectors = 3; 
   
    sparseMat = sparsifySimilarityMatrix(pairMat,kNearestNeighbors);
    [l,lSym,lRW] = computeLaplacianG(sparseMat);
    [spectralEmbedding] = laplacianToEmbedding(lRW,numEigenvectors);
    % The spectralEmbedding is the first few non-trivial eigenvectors of
    % the laplacian. For a connected graph there is a single trivial
    % eigenvector corresponding to an eigenvector of zero.

% Cluster

    clusterEvalFunc = 'DaviesBouldin';
    clusterMethod = 'linkage'; %kmeans, linkage
    nClust = [2:6]; % Number of clusters to try
    % This requires the Matlab stat toolbox.
    evalRes = evalclusters(spectralEmbedding(:,2:3),clusterMethod,clusterEvalFunc,'klist',nClust);
    
% Plot

    ax = gscatter(spectralEmbedding(:,2),spectralEmbedding(:,3),evalRes.OptimalY);