基于单分类器的高分辨率遥感影像道路提取

本人硕士阶段做了很久的高分辨率遥感影像道路提取，颇有心得，在此，本人将最新的研究成果进行开源。。。

大家都知道，传统的基于机器学习的分类方法通常需要正负样本的同时参与，才能得到目标类，但是负样本的勾选，通常很困难，也非常难获得，根据文献-

《Elkan, Charles, and Keith Noto. "Learning classifiers from only positive and unlabeled data." Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2008.》中提出的方法，本人构建了一种单样本分类器，其中原理如下:

根据上面原理，本人基于传统的随机森林算法，构建了单分类器，random forest positive and unlabeled learning classifer，matlab 代码如下所示：

这是主要的单分类器函数：

function Pro= RF_pul(sample,Image,train,r_pul,ntree,depth)
%positive and unlabeled learning ,finished by wang zhipan, southwest
%jiaotonguniversiy;

% sample:the number of train samples,the first column is column in original
% image,the second column is row in original image

% Image:the original image

% train :the ratio of trained sample,
% r_pul: the ratio between positive samples and unlabed samles

% ntree:default 200;
% depth : default 2;

if nargin<3
    train=0.75;
    r_pul=1;
    ntree=200;
    depth=2;
elseif nargin<4
    r_pul=1;
    ntree=200;
    depth=2;
elseif nargin<5
    ntree=200;
    depth=2;
elseif nargin<6
    depth=2;
end

[m_train,n_train]=size(sample);
n_train=n_train-2;

[m,n,~]=size(Image);

for i=1:n_train
    predict(:,i)=reshape(Image(:,:,i),m*n,1);
end

%% train matrix
rand_negtive=randperm(m*n)‘;
n_tr=floor(m_train*train+(m_train*train*r_pul));

Train_matrix=zeros(n_tr,n_train); % the matirx used to train
Label_train=zeros(n_tr,1);  %label
Label_train(1:m_train*train,1)=1;
Label_train(m_train*train+1:end,1)=0;

Train_matrix(1:m_train*train,1:end)=sample(1:m_train*train,3:end);

for i=1:m_train*train*r_pul
    Train_matrix(floor(m_train*train)+i,:)=predict(rand_negtive(i),:);
end

%% predictmatirx
min_val=min(predict);
max_val=max(predict);

for i=1:n_train   %linear normlized
    predict(:,i)=(predict(:,i)-min_val(i))./(max_val(i)-min_val(i));
    Train_matrix(:,i)=(Train_matrix(:,i)-min_val(i))./(max_val(i)-min_val(i));
end

model = classRF_train(Train_matrix,Label_train,ntree,depth);

[~,votes] = classRF_predict(predict,model);
g=votes./ntree;  %use the number of tree the stand for probability

G=reshape(g(:,2),m,n);

valid_index=sub2ind([m,n],sample(m_train*train+1:end,2),sample(m_train*train+1:end,1));

c=mean(G(valid_index));  % the c in paper,we use mean value of all data

Pro=mat2gray(G./c);  %Normalized Probability
end

下面写一个script，进行运行：

%% test RF_pul
clear
tic

Text=textread(‘Sample3.txt‘);
[Img,ref]=geotiffread(‘test3.tif‘);% 获取坐标信息
Img=double(Img);
info=geotiffinfo(‘test3.tif‘);
[m,n,z]=size(Img);

Sample=zeros(size(Text,1),5);
Sample(1:end,1:2)=Text(1:end,2:3);  %以行列的形式存储原始影像坐标,其中第一列表示原始影像中的列，第二列表示原始影像中的行
Sample(1:end,3:end)=Text(1:end,8:end); % 存储特征
clear Text

Pro= RF_pul(Sample,Img);
Proimg=reshape(Pro,m,n);
clear Pro Img
figure,imshow(Proimg),title(‘后验概率‘);

geotiffwrite(‘Probility.tif‘,Proimg,ref,‘GeoKeyDirectoryTag‘, info.GeoTIFFTags.GeoKeyDirectoryTag);

toc

　　得到的最终结果如下所示：

尽请各位同行批评指正，如果对代码有比较难理解的地方，欢迎联系，手机：13094445837，qq1044625113