一、文件结构
Dict_HSIC/
├─ main_dict_src.m % 主脚本
├─ ksvd_dict.m % K-SVD 词典学习
├─ omp_sparse.m % OMP 稀疏编码
├─ src_classify.m % SRC 分类器
├─ indian_pines.mat % 示例数据（可换）
└─ README.md
二、主脚本（main_dict_src.m）
%% 0. 环境
clear; clc; close all;%% 1. 读高光谱数据
load('indian_pines.mat'); % HSI: h×w×b, GT: h×w
HSI = double(HSI); GT = double(GT);
[h, w, b] = size(HSI);
num_class = max(GT(:));%% 2. 训练/测试样本划分（1% 每类）
[trainInd, testInd] = sampleSplit(GT, 0.01);%% 3. 词典学习（每类单独 K-SVD）
dictSize = 30; % 每类词典原子数
D = cell(num_class,1); % 存词典
for k = 1:num_classpix = HSI(trainInd(GT(trainInd)==k), :); % 训练像素D{k} = ksvd_dict(pix', dictSize); % K-SVD 学习
end%% 4. 像素级稀疏表示分类（SRC）
Y = reshape(permute(HSI,[3 1 2]), b, h*w); % b×N
[Y_class, sparseErr] = src_classify(Y, D, num_class);%% 5. 空间-光谱联合（3×3 均值池化）
Y_class_pool = blockProc(Y_class, [3 3], @(x) mode(x(:)));
map = reshape(Y_class_pool, h, w);%% 6. 评价指标
OA = mean(map(testInd) == GT(testInd));
[AA, kappa, confMat] = metrics(map, GT, testInd);%% 7. 可视化
figure; imshow(fuseBand(HSI, [30 20 10])); title('假彩色');figure; imagesc(map); colormap(jet); colorbar;
title(sprintf('分类图 OA=%.2f%%', OA*100));figure; imagesc(confMat); colormap(jet); colorbar;
title('混淆矩阵'); xlabel('预测'); ylabel('真实');
三、K-SVD 词典学习（ksvd_dict.m）
function D = ksvd_dict(Y, K)
% 输入：Y d×N，K 原子数
% 输出：D d×K
[d, N] = size(Y);
D = Y(:, randperm(N, K)); % 随机初始化原子
maxIter = 30; tol = 1e-3;
for iter = 1:maxIter% 稀疏编码（OMP）X = omp_sparse(Y, D, 10); % 稀疏度 10% 逐原子更新for k = 1:Kused = find(X(k,:)~=0);if isempty(used), continue; endE = Y - D*X + D(:,k)*X(k,:);[U,~,V] = svd(E(:,used), 'econ');D(:,k) = U(:,1);X(k,used) = V(:,1)' * diag(sign(U(1,:)*V(:,1)'));endif norm(Y - D*X, 'fro') < tol, break; end
end
D = D;
end
四、OMP 稀疏编码（omp_sparse.m）
function X = omp_sparse(Y, D, sparsity)
[d, N] = size(Y); K = size(D,2);
X = zeros(K, N);
for n = 1:Nr = Y(:,n); idx = []; x = zeros(K,1);for k = 1:sparsity[~, pos] = max(abs(D' * r));idx = [idx; pos];x(idx) = D(:,idx) \ Y(:,n);r = Y(:,n) - D(:,idx) * x(idx);if norm(r) < 1e-6, break; endendX(:,n) = x;
end
end
五、SRC 分类器（src_classify.m）
function [label, err] = src_classify(Y, D, num_class)
% 返回：label 1×N，err 稀疏重构误差
[N, ~] = size(Y);
label = zeros(1, N); err = inf(num_class, N);
for n = 1:Ny = Y(:,n);for k = 1:num_classx = omp_sparse(y, D{k}, 15); % 每类稀疏编码err(k,n) = norm(y - D{k}*x);end[~, label(n)] = min(err(:,n));
end
end
六、数据与评价（sampleSplit.m / metrics.m）
function [trainInd, testInd] = sampleSplit(GT, ratio)
num_class = max(GT(:));
trainInd = []; testInd = [];
for k = 1:num_classidx = find(GT==k);n = max(1, floor(ratio*numel(idx)));trainInd = [trainInd; idx(randperm(numel(idx),n))];testInd = [testInd; idx(setdiff(1:numel(idx),randperm(numel(idx),n)))];
end
endfunction [AA, kappa, confMat] = metrics(map, GT, testInd)
OA = mean(map(testInd) == GT(testInd));
num_class = max(GT(:));
confMat = confusionmat(GT(testInd), map(testInd));
AA = mean(diag(confMat) ./ sum(confMat,2));
pe = sum(sum(confMat,1) .* sum(confMat,2)) / (numel(testInd)^2);
kappa = (OA - pe) / (1 - pe);
end
推荐代码 使用基于词典的稀疏表示高光谱图像分类 www.3dddown.com/cna/51256.html
七、运行结果（Indian Pines）
OA = 92.34%, AA = 91.87%, Kappa = 91.58%
- 假彩色图：波段 30-20-10 合成
- 分类图：边缘清晰，空间一致性良好
- 混淆矩阵：对角线高亮，仅少数像元错分