24 #include "imageclasses/ImgReaderGdal.h"
25 #include "imageclasses/ImgWriterGdal.h"
26 #include "imageclasses/ImgReaderOgr.h"
27 #include "imageclasses/ImgWriterOgr.h"
28 #include "base/Optionpk.h"
29 #include "base/PosValue.h"
30 #include "algorithms/ConfusionMatrix.h"
31 #include "algorithms/svm.h"
38 enum SVM_TYPE {C_SVC=0, nu_SVC=1,one_class=2, epsilon_SVR=3, nu_SVR=4};
39 enum KERNEL_TYPE {linear=0,polynomial=1,radial=2,sigmoid=3};
42 #define Malloc(type,n) (type *)malloc((n)*sizeof(type))
46 int main(
int argc,
char *argv[])
48 vector<double> priors;
52 Optionpk<string> training_opt(
"t",
"training",
"Training vector file. A single vector file contains all training features (must be set as: b0, b1, b2,...) for all classes (class numbers identified by label option). Use multiple training files for bootstrap aggregation (alternative to the bag and bsize options, where a random subset is taken from a single training file)");
54 Optionpk<string> label_opt(
"label",
"label",
"Attribute name for class label in training vector file.",
"label");
55 Optionpk<unsigned int> balance_opt(
"bal",
"balance",
"Balance the input data to this number of samples for each class", 0);
56 Optionpk<bool> random_opt(
"random",
"random",
"Randomize training data for balancing and bagging",
true, 2);
57 Optionpk<int> minSize_opt(
"min",
"min",
"If number of training pixels is less then min, do not take this class into account (0: consider all classes)", 0);
58 Optionpk<short> band_opt(
"b",
"band",
"Band index (starting from 0, either use band option or use start to end)");
60 Optionpk<double> bend_opt(
"e",
"end",
"End band sequence number (set to 0 to include all bands)", 0);
61 Optionpk<double> offset_opt(
"\0",
"offset",
"Offset value for each spectral band input features: refl[band]=(DN[band]-offset[band])/scale[band]", 0.0);
62 Optionpk<double> scale_opt(
"\0",
"scale",
"Scale value for each spectral band input features: refl=(DN[band]-offset[band])/scale[band] (use 0 if scale min and max in each band to -1.0 and 1.0)", 0.0);
63 Optionpk<double> priors_opt(
"prior",
"prior",
"Prior probabilities for each class (e.g., -p 0.3 -p 0.3 -p 0.2 ). Used for input only (ignored for cross validation)", 0.0);
64 Optionpk<string> priorimg_opt(
"pim",
"priorimg",
"Prior probability image (multi-band img with band for each class",
"",2);
66 Optionpk<std::string> svm_type_opt(
"svmt",
"svmtype",
"Type of SVM (C_SVC, nu_SVC,one_class, epsilon_SVR, nu_SVR)",
"C_SVC");
67 Optionpk<std::string> kernel_type_opt(
"kt",
"kerneltype",
"Type of kernel function (linear,polynomial,radial,sigmoid) ",
"radial");
69 Optionpk<float> gamma_opt(
"g",
"gamma",
"Gamma in kernel function",1.0);
71 Optionpk<float> ccost_opt(
"cc",
"ccost",
"The parameter C of C_SVC, epsilon_SVR, and nu_SVR",1000);
72 Optionpk<float> nu_opt(
"nu",
"nu",
"The parameter nu of nu_SVC, one_class SVM, and nu_SVR",0.5);
73 Optionpk<float> epsilon_loss_opt(
"eloss",
"eloss",
"The epsilon in loss function of epsilon_SVR",0.1);
74 Optionpk<int> cache_opt(
"cache",
"cache",
"Cache memory size in MB",100);
75 Optionpk<float> epsilon_tol_opt(
"etol",
"etol",
"The tolerance of termination criterion",0.001);
76 Optionpk<bool> shrinking_opt(
"shrink",
"shrink",
"Whether to use the shrinking heuristics",
false);
77 Optionpk<bool> prob_est_opt(
"pe",
"probest",
"Whether to train a SVC or SVR model for probability estimates",
true,2);
79 Optionpk<unsigned short> comb_opt(
"comb",
"comb",
"How to combine bootstrap aggregation classifiers (0: sum rule, 1: product rule, 2: max rule). Also used to aggregate classes with rc option.",0);
81 Optionpk<int> bagSize_opt(
"bagsize",
"bagsize",
"Percentage of features used from available training features for each bootstrap aggregation (one size for all classes, or a different size for each class respectively", 100);
82 Optionpk<string> classBag_opt(
"cb",
"classbag",
"Output for each individual bootstrap aggregation");
83 Optionpk<string> mask_opt(
"m",
"mask",
"Use the first band of the specified file as a validity mask. Nodata values can be set with the option msknodata.");
84 Optionpk<short> msknodata_opt(
"msknodata",
"msknodata",
"Mask value(s) not to consider for classification (use negative values if only these values should be taken into account). Values will be taken over in classification image.", 0);
87 Optionpk<string> oformat_opt(
"of",
"oformat",
"Output image format (see also gdal_translate). Empty string: inherit from input image");
88 Optionpk<string> option_opt(
"co",
"co",
"Creation option for output file. Multiple options can be specified.");
89 Optionpk<string> colorTable_opt(
"ct",
"ct",
"Color table in ASCII format having 5 columns: id R G B ALFA (0: transparent, 255: solid)");
91 Optionpk<string> entropy_opt(
"entropy",
"entropy",
"Entropy image (measure for uncertainty of classifier output",
"",2);
92 Optionpk<string> active_opt(
"active",
"active",
"Ogr output for active training sample.",
"",2);
93 Optionpk<string> ogrformat_opt(
"f",
"f",
"Output ogr format for active training sample",
"SQLite");
96 Optionpk<short> classvalue_opt(
"r",
"reclass",
"List of class values (use same order as in class opt).");
100 bstart_opt.setHide(1);
102 balance_opt.setHide(1);
103 minSize_opt.setHide(1);
105 bagSize_opt.setHide(1);
107 classBag_opt.setHide(1);
109 priorimg_opt.setHide(1);
110 offset_opt.setHide(1);
111 scale_opt.setHide(1);
112 svm_type_opt.setHide(1);
113 kernel_type_opt.setHide(1);
114 kernel_degree_opt.setHide(1);
115 coef0_opt.setHide(1);
117 epsilon_loss_opt.setHide(1);
118 cache_opt.setHide(1);
119 epsilon_tol_opt.setHide(1);
120 shrinking_opt.setHide(1);
121 prob_est_opt.setHide(1);
122 entropy_opt.setHide(1);
123 active_opt.setHide(1);
124 nactive_opt.setHide(1);
125 verbose_opt.setHide(1);
126 random_opt.setHide(1);
130 doProcess=training_opt.retrieveOption(argc,argv);
131 input_opt.retrieveOption(argc,argv);
132 output_opt.retrieveOption(argc,argv);
133 cv_opt.retrieveOption(argc,argv);
134 tlayer_opt.retrieveOption(argc,argv);
135 classname_opt.retrieveOption(argc,argv);
136 classvalue_opt.retrieveOption(argc,argv);
137 oformat_opt.retrieveOption(argc,argv);
138 ogrformat_opt.retrieveOption(argc,argv);
139 option_opt.retrieveOption(argc,argv);
140 colorTable_opt.retrieveOption(argc,argv);
141 label_opt.retrieveOption(argc,argv);
142 priors_opt.retrieveOption(argc,argv);
143 gamma_opt.retrieveOption(argc,argv);
144 ccost_opt.retrieveOption(argc,argv);
145 mask_opt.retrieveOption(argc,argv);
146 msknodata_opt.retrieveOption(argc,argv);
147 nodata_opt.retrieveOption(argc,argv);
149 band_opt.retrieveOption(argc,argv);
150 bstart_opt.retrieveOption(argc,argv);
151 bend_opt.retrieveOption(argc,argv);
152 balance_opt.retrieveOption(argc,argv);
153 minSize_opt.retrieveOption(argc,argv);
154 bag_opt.retrieveOption(argc,argv);
155 bagSize_opt.retrieveOption(argc,argv);
156 comb_opt.retrieveOption(argc,argv);
157 classBag_opt.retrieveOption(argc,argv);
158 prob_opt.retrieveOption(argc,argv);
159 priorimg_opt.retrieveOption(argc,argv);
160 offset_opt.retrieveOption(argc,argv);
161 scale_opt.retrieveOption(argc,argv);
162 svm_type_opt.retrieveOption(argc,argv);
163 kernel_type_opt.retrieveOption(argc,argv);
164 kernel_degree_opt.retrieveOption(argc,argv);
165 coef0_opt.retrieveOption(argc,argv);
166 nu_opt.retrieveOption(argc,argv);
167 epsilon_loss_opt.retrieveOption(argc,argv);
168 cache_opt.retrieveOption(argc,argv);
169 epsilon_tol_opt.retrieveOption(argc,argv);
170 shrinking_opt.retrieveOption(argc,argv);
171 prob_est_opt.retrieveOption(argc,argv);
172 entropy_opt.retrieveOption(argc,argv);
173 active_opt.retrieveOption(argc,argv);
174 nactive_opt.retrieveOption(argc,argv);
175 verbose_opt.retrieveOption(argc,argv);
176 random_opt.retrieveOption(argc,argv);
178 catch(
string predefinedString){
179 std::cout << predefinedString << std::endl;
184 cout <<
"Usage: pksvm -t training [-i input -o output] [-cv value]" << endl;
186 std::cout <<
"short option -h shows basic options only, use long option --help to show all options" << std::endl;
190 if(entropy_opt[0]==
"")
192 if(active_opt[0]==
"")
194 if(priorimg_opt[0]==
"")
195 priorimg_opt.clear();
198 std::map<std::string, svm::SVM_TYPE> svmMap;
200 svmMap[
"C_SVC"]=svm::C_SVC;
201 svmMap[
"nu_SVC"]=svm::nu_SVC;
202 svmMap[
"one_class"]=svm::one_class;
203 svmMap[
"epsilon_SVR"]=svm::epsilon_SVR;
204 svmMap[
"nu_SVR"]=svm::nu_SVR;
206 std::map<std::string, svm::KERNEL_TYPE> kernelMap;
208 kernelMap[
"linear"]=svm::linear;
209 kernelMap[
"polynomial"]=svm::polynomial;
210 kernelMap[
"radial"]=svm::radial;
211 kernelMap[
"sigmoid;"]=svm::sigmoid;
213 assert(training_opt.size());
215 if(verbose_opt[0]>=1){
217 std::cout <<
"input filename: " << input_opt[0] << std::endl;
219 std::cout <<
"mask filename: " << mask_opt[0] << std::endl;
220 std::cout <<
"training vector file: " << std::endl;
221 for(
int ifile=0;ifile<training_opt.size();++ifile)
222 std::cout << training_opt[ifile] << std::endl;
223 std::cout <<
"verbose: " << verbose_opt[0] << std::endl;
225 unsigned short nbag=(training_opt.size()>1)?training_opt.size():bag_opt[0];
226 if(verbose_opt[0]>=1)
227 std::cout <<
"number of bootstrap aggregations: " << nbag << std::endl;
230 if(active_opt.size()){
231 prob_est_opt[0]=
true;
233 activeWriter.open(active_opt[0],ogrformat_opt[0]);
234 activeWriter.createLayer(active_opt[0],trainingReader.getProjection(),wkbPoint,NULL);
235 activeWriter.copyFields(trainingReader);
237 vector<PosValue> activePoints(nactive_opt[0]);
238 for(
int iactive=0;iactive<activePoints.size();++iactive){
239 activePoints[iactive].value=1.0;
240 activePoints[iactive].posx=0.0;
241 activePoints[iactive].posy=0.0;
244 unsigned int totalSamples=0;
245 unsigned int nactive=0;
246 vector<struct svm_model*> svm(nbag);
247 vector<struct svm_parameter> param(nbag);
254 if(priors_opt.size()>1){
255 priors.resize(priors_opt.size());
257 for(
short iclass=0;iclass<priors_opt.size();++iclass){
258 priors[iclass]=priors_opt[iclass];
259 normPrior+=priors[iclass];
262 for(
short iclass=0;iclass<priors_opt.size();++iclass)
263 priors[iclass]/=normPrior;
268 std::sort(band_opt.begin(),band_opt.end());
270 map<string,short> classValueMap;
271 vector<std::string> nameVector;
272 if(classname_opt.size()){
273 assert(classname_opt.size()==classvalue_opt.size());
274 for(
int iclass=0;iclass<classname_opt.size();++iclass)
275 classValueMap[classname_opt[iclass]]=classvalue_opt[iclass];
280 vector< vector<double> > offset(nbag);
281 vector< vector<double> > scale(nbag);
282 map<string,Vector2d<float> > trainingMap;
283 vector< Vector2d<float> > trainingPixels;
284 vector<string> fields;
286 vector<struct svm_problem> prob(nbag);
287 vector<struct svm_node *> x_space(nbag);
289 for(
int ibag=0;ibag<nbag;++ibag){
291 if(ibag<training_opt.size()){
293 trainingPixels.clear();
294 if(verbose_opt[0]>=1)
295 std::cout <<
"reading imageVector file " << training_opt[0] << std::endl;
299 totalSamples=trainingReaderBag.readDataImageOgr(trainingMap,fields,band_opt,label_opt[0],tlayer_opt,verbose_opt[0]);
301 totalSamples=trainingReaderBag.readDataImageOgr(trainingMap,fields,bstart_opt[0],bend_opt[0],label_opt[0],tlayer_opt,verbose_opt[0]);
302 if(trainingMap.size()<2){
303 string errorstring=
"Error: could not read at least two classes from training file, did you provide class labels in training sample (see option label)?";
306 trainingReaderBag.close();
309 cerr << error << std::endl;
312 catch(std::exception& e){
313 std::cerr <<
"Error: ";
314 std::cerr << e.what() << std::endl;
315 std::cerr << CPLGetLastErrorMsg() << std::endl;
319 cerr <<
"error catched" << std::endl;
326 std::cout <<
"training pixels: " << std::endl;
327 map<string,Vector2d<float> >::iterator mapit=trainingMap.begin();
328 while(mapit!=trainingMap.end()){
330 if((mapit->second).size()<minSize_opt[0]){
331 trainingMap.erase(mapit);
334 trainingPixels.push_back(mapit->second);
336 std::cout << mapit->first <<
": " << (mapit->second).size() <<
" samples" << std::endl;
340 nclass=trainingPixels.size();
341 if(classname_opt.size())
342 assert(nclass==classname_opt.size());
343 nband=trainingPixels[0][0].size()-2;
346 assert(nclass==trainingPixels.size());
347 assert(nband==trainingPixels[0][0].size()-2);
352 if(balance_opt[0]>0){
353 while(balance_opt.size()<nclass)
354 balance_opt.push_back(balance_opt.back());
358 for(
short iclass=0;iclass<nclass;++iclass){
359 if(trainingPixels[iclass].size()>balance_opt[iclass]){
360 while(trainingPixels[iclass].size()>balance_opt[iclass]){
361 int index=rand()%trainingPixels[iclass].size();
362 trainingPixels[iclass].erase(trainingPixels[iclass].begin()+index);
366 int oldsize=trainingPixels[iclass].size();
367 for(
int isample=trainingPixels[iclass].size();isample<balance_opt[iclass];++isample){
368 int index = rand()%oldsize;
369 trainingPixels[iclass].push_back(trainingPixels[iclass][index]);
372 totalSamples+=trainingPixels[iclass].size();
377 offset[ibag].resize(nband);
378 scale[ibag].resize(nband);
379 if(offset_opt.size()>1)
380 assert(offset_opt.size()==nband);
381 if(scale_opt.size()>1)
382 assert(scale_opt.size()==nband);
383 for(
int iband=0;iband<nband;++iband){
384 if(verbose_opt[0]>=1)
385 std::cout <<
"scaling for band" << iband << std::endl;
386 offset[ibag][iband]=(offset_opt.size()==1)?offset_opt[0]:offset_opt[iband];
387 scale[ibag][iband]=(scale_opt.size()==1)?scale_opt[0]:scale_opt[iband];
389 if(scale[ibag][iband]<=0){
390 float theMin=trainingPixels[0][0][iband+startBand];
391 float theMax=trainingPixels[0][0][iband+startBand];
392 for(
short iclass=0;iclass<nclass;++iclass){
393 for(
int isample=0;isample<trainingPixels[iclass].size();++isample){
394 if(theMin>trainingPixels[iclass][isample][iband+startBand])
395 theMin=trainingPixels[iclass][isample][iband+startBand];
396 if(theMax<trainingPixels[iclass][isample][iband+startBand])
397 theMax=trainingPixels[iclass][isample][iband+startBand];
400 offset[ibag][iband]=theMin+(theMax-theMin)/2.0;
401 scale[ibag][iband]=(theMax-theMin)/2.0;
402 if(verbose_opt[0]>=1){
403 std::cout <<
"Extreme image values for band " << iband <<
": [" << theMin <<
"," << theMax <<
"]" << std::endl;
404 std::cout <<
"Using offset, scale: " << offset[ibag][iband] <<
", " << scale[ibag][iband] << std::endl;
405 std::cout <<
"scaled values for band " << iband <<
": [" << (theMin-offset[ibag][iband])/scale[ibag][iband] <<
"," << (theMax-offset[ibag][iband])/scale[ibag][iband] <<
"]" << std::endl;
411 offset[ibag].resize(nband);
412 scale[ibag].resize(nband);
413 for(
int iband=0;iband<nband;++iband){
414 offset[ibag][iband]=offset[0][iband];
415 scale[ibag][iband]=scale[0][iband];
420 if(priors_opt.size()==1){
421 priors.resize(nclass);
422 for(
short iclass=0;iclass<nclass;++iclass)
423 priors[iclass]=1.0/nclass;
425 assert(priors_opt.size()==1||priors_opt.size()==nclass);
428 while(bagSize_opt.size()<nclass)
429 bagSize_opt.push_back(bagSize_opt.back());
431 if(verbose_opt[0]>=1){
432 std::cout <<
"number of bands: " << nband << std::endl;
433 std::cout <<
"number of classes: " << nclass << std::endl;
434 if(priorimg_opt.empty()){
435 std::cout <<
"priors:";
436 for(
short iclass=0;iclass<nclass;++iclass)
437 std::cout <<
" " << priors[iclass];
438 std::cout << std::endl;
441 map<string,Vector2d<float> >::iterator mapit=trainingMap.begin();
444 while(mapit!=trainingMap.end()){
445 nameVector.push_back(mapit->first);
446 if(classValueMap.size()){
448 if(classValueMap[mapit->first]>0){
449 if(cm.getClassIndex(type2string<short>(classValueMap[mapit->first]))<0){
450 cm.pushBackClassName(type2string<short>(classValueMap[mapit->first]),doSort);
454 std::cerr <<
"Error: names in classname option are not complete, please check names in training vector and make sure classvalue is > 0" << std::endl;
459 cm.pushBackClassName(mapit->first,doSort);
464 std::cerr <<
"Error: did you provide class pairs names (-c) and integer values (-r) for each class in training vector?" << std::endl;
467 if(classname_opt.empty()){
469 for(
int iclass=0;iclass<nclass;++iclass){
471 std::cout << iclass <<
" " << cm.getClass(iclass) <<
" -> " << string2type<short>(cm.getClass(iclass)) << std::endl;
472 classValueMap[cm.getClass(iclass)]=string2type<short>(cm.getClass(iclass));
484 vector< Vector2d<float> > trainingFeatures(nclass);
485 for(
short iclass=0;iclass<nclass;++iclass){
487 if(verbose_opt[0]>=1)
488 std::cout <<
"calculating features for class " << iclass << std::endl;
491 nctraining=(bagSize_opt[iclass]<100)? trainingPixels[iclass].size()/100.0*bagSize_opt[iclass] : trainingPixels[iclass].size();
494 assert(nctraining<=trainingPixels[iclass].size());
496 if(bagSize_opt[iclass]<100)
497 random_shuffle(trainingPixels[iclass].begin(),trainingPixels[iclass].end());
499 std::cout <<
"nctraining (class " << iclass <<
"): " << nctraining << std::endl;
500 trainingFeatures[iclass].resize(nctraining);
501 for(
int isample=0;isample<nctraining;++isample){
503 for(
int iband=0;iband<nband;++iband){
504 float value=trainingPixels[iclass][isample][iband+startBand];
505 trainingFeatures[iclass][isample].push_back((value-offset[ibag][iband])/scale[ibag][iband]);
508 assert(trainingFeatures[iclass].size()==nctraining);
511 unsigned int nFeatures=trainingFeatures[0][0].size();
512 if(verbose_opt[0]>=1)
513 std::cout <<
"number of features: " << nFeatures << std::endl;
514 unsigned int ntraining=0;
515 for(
short iclass=0;iclass<nclass;++iclass)
516 ntraining+=trainingFeatures[iclass].size();
517 if(verbose_opt[0]>=1)
518 std::cout <<
"training size over all classes: " << ntraining << std::endl;
520 prob[ibag].l=ntraining;
521 prob[ibag].y = Malloc(
double,prob[ibag].l);
522 prob[ibag].x = Malloc(
struct svm_node *,prob[ibag].l);
523 x_space[ibag] = Malloc(
struct svm_node,(nFeatures+1)*ntraining);
524 unsigned long int spaceIndex=0;
526 for(
short iclass=0;iclass<nclass;++iclass){
527 for(
int isample=0;isample<trainingFeatures[iclass].size();++isample){
528 prob[ibag].x[lIndex]=&(x_space[ibag][spaceIndex]);
529 for(
int ifeature=0;ifeature<nFeatures;++ifeature){
530 x_space[ibag][spaceIndex].index=ifeature+1;
531 x_space[ibag][spaceIndex].value=trainingFeatures[iclass][isample][ifeature];
534 x_space[ibag][spaceIndex++].index=-1;
535 prob[ibag].y[lIndex]=iclass;
539 assert(lIndex==prob[ibag].l);
542 param[ibag].svm_type = svmMap[svm_type_opt[0]];
543 param[ibag].kernel_type = kernelMap[kernel_type_opt[0]];
544 param[ibag].degree = kernel_degree_opt[0];
545 param[ibag].gamma = (gamma_opt[0]>0)? gamma_opt[0] : 1.0/nFeatures;
546 param[ibag].coef0 = coef0_opt[0];
547 param[ibag].nu = nu_opt[0];
548 param[ibag].cache_size = cache_opt[0];
549 param[ibag].C = ccost_opt[0];
550 param[ibag].eps = epsilon_tol_opt[0];
551 param[ibag].p = epsilon_loss_opt[0];
552 param[ibag].shrinking = (shrinking_opt[0])? 1 : 0;
553 param[ibag].probability = (prob_est_opt[0])? 1 : 0;
554 param[ibag].nr_weight = 0;
555 param[ibag].weight_label = NULL;
556 param[ibag].weight = NULL;
557 param[ibag].verbose=(verbose_opt[0]>1)?
true:
false;
560 std::cout <<
"checking parameters" << std::endl;
561 svm_check_parameter(&prob[ibag],¶m[ibag]);
563 std::cout <<
"parameters ok, training" << std::endl;
564 svm[ibag]=svm_train(&prob[ibag],¶m[ibag]);
566 std::cout <<
"SVM is now trained" << std::endl;
569 std::cout <<
"Cross validating" << std::endl;
570 double *target = Malloc(
double,prob[ibag].l);
571 svm_cross_validation(&prob[ibag],¶m[ibag],cv_opt[0],target);
572 assert(param[ibag].svm_type != EPSILON_SVR&¶m[ibag].svm_type != NU_SVR);
574 for(
int i=0;i<prob[ibag].l;i++){
575 string refClassName=nameVector[prob[ibag].y[i]];
576 string className=nameVector[target[i]];
577 if(classValueMap.size())
578 cm.incrementResult(type2string<short>(classValueMap[refClassName]),type2string<short>(classValueMap[className]),1.0/nbag);
580 cm.incrementResult(cm.getClass(prob[ibag].y[i]),cm.getClass(target[i]),1.0/nbag);
589 assert(cm.nReference());
590 std::cout << cm << std::endl;
591 cout <<
"class #samples userAcc prodAcc" << endl;
598 for(
short iclass=0;iclass<cm.nClasses();++iclass){
599 dua=cm.ua(cm.getClass(iclass),&se95_ua);
600 dpa=cm.pa(cm.getClass(iclass),&se95_pa);
601 cout << cm.getClass(iclass) <<
" " << cm.nReference(cm.getClass(iclass)) <<
" " << dua <<
" (" << se95_ua <<
")" <<
" " << dpa <<
" (" << se95_pa <<
")" << endl;
603 std::cout <<
"Kappa: " << cm.kappa() << std::endl;
605 std::cout <<
"Overall Accuracy: " << 100*doa <<
" (" << 100*se95_oa <<
")" << std::endl;
609 if(input_opt.empty())
612 const char* pszMessage;
613 void* pProgressArg=NULL;
614 GDALProgressFunc pfnProgress=GDALTermProgress;
617 pfnProgress(progress,pszMessage,pProgressArg);
619 bool inputIsRaster=
false;
622 imgReaderOgr.open(input_opt[0]);
623 imgReaderOgr.close();
625 catch(
string errorString){
631 if(verbose_opt[0]>=1)
632 std::cout <<
"opening image " << input_opt[0] << std::endl;
633 testImage.open(input_opt[0]);
636 cerr << error << std::endl;
642 if(verbose_opt[0]>=1)
643 std::cout <<
"opening mask image file " << mask_opt[0] << std::endl;
644 maskReader.open(mask_opt[0]);
647 cerr << error << std::endl;
651 cerr <<
"error catched" << std::endl;
656 if(priorimg_opt.size()){
658 if(verbose_opt[0]>=1)
659 std::cout <<
"opening prior image " << priorimg_opt[0] << std::endl;
660 priorReader.open(priorimg_opt[0]);
661 assert(priorReader.nrOfCol()==testImage.nrOfCol());
662 assert(priorReader.nrOfRow()==testImage.nrOfRow());
665 cerr << error << std::endl;
669 cerr <<
"error catched" << std::endl;
674 int nrow=testImage.nrOfRow();
675 int ncol=testImage.nrOfCol();
676 if(option_opt.findSubstring(
"INTERLEAVE=")==option_opt.end()){
677 string theInterleave=
"INTERLEAVE=";
678 theInterleave+=testImage.getInterleave();
679 option_opt.push_back(theInterleave);
681 vector<char> classOut(ncol);
689 string imageType=testImage.getImageType();
690 if(oformat_opt.size())
691 imageType=oformat_opt[0];
693 assert(output_opt.size());
694 if(verbose_opt[0]>=1)
695 std::cout <<
"opening class image for writing output " << output_opt[0] << std::endl;
696 if(classBag_opt.size()){
697 classImageBag.open(classBag_opt[0],ncol,nrow,nbag,GDT_Byte,imageType,option_opt);
698 classImageBag.GDALSetNoDataValue(nodata_opt[0]);
699 classImageBag.copyGeoTransform(testImage);
700 classImageBag.setProjection(testImage.getProjection());
702 classImageOut.open(output_opt[0],ncol,nrow,1,GDT_Byte,imageType,option_opt);
703 classImageOut.GDALSetNoDataValue(nodata_opt[0]);
704 classImageOut.copyGeoTransform(testImage);
705 classImageOut.setProjection(testImage.getProjection());
706 if(colorTable_opt.size())
707 classImageOut.setColorTable(colorTable_opt[0],0);
709 probImage.open(prob_opt[0],ncol,nrow,nclass,GDT_Byte,imageType,option_opt);
710 probImage.GDALSetNoDataValue(nodata_opt[0]);
711 probImage.copyGeoTransform(testImage);
712 probImage.setProjection(testImage.getProjection());
714 if(entropy_opt.size()){
715 entropyImage.open(entropy_opt[0],ncol,nrow,1,GDT_Byte,imageType,option_opt);
716 entropyImage.GDALSetNoDataValue(nodata_opt[0]);
717 entropyImage.copyGeoTransform(testImage);
718 entropyImage.setProjection(testImage.getProjection());
722 cerr << error << std::endl;
725 for(
int iline=0;iline<nrow;++iline){
726 vector<float> buffer(ncol);
727 vector<short> lineMask;
729 if(priorimg_opt.size())
730 linePrior.resize(nclass,ncol);
733 vector<float> entropy(ncol);
735 if(classBag_opt.size())
736 classBag.resize(nbag,ncol);
739 for(
int iband=0;iband<band_opt.size();++iband){
740 if(verbose_opt[0]==2)
741 std::cout <<
"reading band " << band_opt[iband] << std::endl;
742 assert(band_opt[iband]>=0);
743 assert(band_opt[iband]<testImage.nrOfBand());
744 testImage.readData(buffer,GDT_Float32,iline,band_opt[iband]);
745 for(
int icol=0;icol<ncol;++icol)
746 hpixel[icol].push_back(buffer[icol]);
750 for(
int iband=bstart_opt[0];iband<bstart_opt[0]+nband;++iband){
751 if(verbose_opt[0]==2)
752 std::cout <<
"reading band " << iband << std::endl;
754 assert(iband<testImage.nrOfBand());
755 testImage.readData(buffer,GDT_Float32,iline,iband);
756 for(
int icol=0;icol<ncol;++icol)
757 hpixel[icol].push_back(buffer[icol]);
761 catch(
string theError){
762 cerr <<
"Error reading " << input_opt[0] <<
": " << theError << std::endl;
766 cerr <<
"error catched" << std::endl;
769 assert(nband==hpixel[0].size());
771 std::cout <<
"used bands: " << nband << std::endl;
773 if(priorimg_opt.size()){
775 for(
short iclass=0;iclass<nclass;++iclass){
776 if(verbose_opt.size()>1)
777 std::cout <<
"Reading " << priorimg_opt[0] <<
" band " << iclass <<
" line " << iline << std::endl;
778 priorReader.readData(linePrior[iclass],GDT_Float32,iline,iclass);
781 catch(
string theError){
782 std::cerr <<
"Error reading " << priorimg_opt[0] <<
": " << theError << std::endl;
786 cerr <<
"error catched" << std::endl;
790 double oldRowMask=-1;
792 for(
int icol=0;icol<ncol;++icol){
793 assert(hpixel[icol].size()==nband);
802 testImage.image2geo(icol,iline,geox,geoy);
803 maskReader.geo2image(geox,geoy,colMask,rowMask);
804 colMask=
static_cast<int>(colMask);
805 rowMask=
static_cast<int>(rowMask);
806 if(rowMask>=0&&rowMask<maskReader.nrOfRow()&&colMask>=0&&colMask<maskReader.nrOfCol()){
807 if(static_cast<int>(rowMask)!=
static_cast<int>(oldRowMask)){
808 assert(rowMask>=0&&rowMask<maskReader.nrOfRow());
811 maskReader.readData(lineMask,GDT_Int16,static_cast<int>(rowMask));
813 catch(
string errorstring){
814 cerr << errorstring << endl;
818 cerr <<
"error catched" << std::endl;
824 for(
short ivalue=0;ivalue<msknodata_opt.size();++ivalue){
825 if(msknodata_opt[ivalue]>=0){
826 if(lineMask[colMask]==msknodata_opt[ivalue]){
827 theMask=lineMask[colMask];
833 if(lineMask[colMask]!=-msknodata_opt[ivalue]){
834 theMask=lineMask[colMask];
844 if(classBag_opt.size())
845 for(
int ibag=0;ibag<nbag;++ibag)
846 classBag[ibag][icol]=theMask;
847 classOut[icol]=theMask;
852 for(
int iband=0;iband<hpixel[icol].size();++iband){
853 if(hpixel[icol][iband]){
859 if(classBag_opt.size())
860 for(
int ibag=0;ibag<nbag;++ibag)
861 classBag[ibag][icol]=nodata_opt[0];
862 classOut[icol]=nodata_opt[0];
866 for(
short iclass=0;iclass<nclass;++iclass)
867 probOut[iclass][icol]=0;
869 std::cout <<
"begin classification " << std::endl;
871 for(
int ibag=0;ibag<nbag;++ibag){
872 vector<double> result(nclass);
875 for(
int iband=0;iband<nband;++iband){
876 x[iband].index=iband+1;
877 x[iband].value=(hpixel[icol][iband]-offset[ibag][iband])/scale[ibag][iband];
880 double predict_label=0;
881 vector<float> prValues(nclass);
883 if(!prob_est_opt[0]){
884 predict_label = svm_predict(svm[ibag],x);
885 for(
short iclass=0;iclass<nclass;++iclass){
886 if(iclass==static_cast<short>(predict_label))
893 assert(svm_check_probability_model(svm[ibag]));
894 predict_label = svm_predict_probability(svm[ibag],x,&(result[0]));
897 if(classBag_opt.size()){
900 classBag[ibag][icol]=0;
903 if(priorimg_opt.size()){
904 for(
short iclass=0;iclass<nclass;++iclass)
905 normPrior+=linePrior[iclass][icol];
907 for(
short iclass=0;iclass<nclass;++iclass){
908 if(priorimg_opt.size())
909 priors[iclass]=linePrior[iclass][icol]/normPrior;
913 probOut[iclass][icol]+=result[iclass]*priors[iclass];
916 probOut[iclass][icol]*=pow(static_cast<float>(priors[iclass]),static_cast<float>(1.0-nbag)/nbag)*result[iclass];
919 if(priors[iclass]*result[iclass]>probOut[iclass][icol])
920 probOut[iclass][icol]=priors[iclass]*result[iclass];
923 if(classBag_opt.size()){
929 if(result[iclass]>maxP){
931 classBag[ibag][icol]=iclass;
942 for(
short iclass=0;iclass<nclass;++iclass){
943 if(probOut[iclass][icol]>maxBag1){
944 maxBag1=probOut[iclass][icol];
945 classOut[icol]=classValueMap[nameVector[iclass]];
947 else if(probOut[iclass][icol]>maxBag2)
948 maxBag2=probOut[iclass][icol];
949 normBag+=probOut[iclass][icol];
953 for(
short iclass=0;iclass<nclass;++iclass){
954 float prv=probOut[iclass][icol];
956 entropy[icol]-=prv*log(prv)/log(2.0);
959 probOut[iclass][icol]=
static_cast<short>(prv+0.5);
964 entropy[icol]/=log(static_cast<double>(nclass))/log(2.0);
965 entropy[icol]=
static_cast<short>(100*entropy[icol]+0.5);
966 if(active_opt.size()){
967 if(entropy[icol]>activePoints.back().value){
968 activePoints.back().value=entropy[icol];
969 activePoints.back().posx=icol;
970 activePoints.back().posy=iline;
973 std::cout << activePoints.back().posx <<
" " << activePoints.back().posy <<
" " << activePoints.back().value << std::endl;
978 if(classBag_opt.size())
979 for(
int ibag=0;ibag<nbag;++ibag)
980 classImageBag.writeData(classBag[ibag],GDT_Byte,iline,ibag);
982 for(
short iclass=0;iclass<nclass;++iclass)
983 probImage.writeData(probOut[iclass],GDT_Float32,iline,iclass);
985 if(entropy_opt.size()){
986 entropyImage.writeData(entropy,GDT_Float32,iline);
988 classImageOut.writeData(classOut,GDT_Byte,iline);
990 progress=
static_cast<float>(iline+1.0)/classImageOut.nrOfRow();
991 pfnProgress(progress,pszMessage,pProgressArg);
995 if(active_opt.size()){
996 for(
int iactive=0;iactive<activePoints.size();++iactive){
997 std::map<string,double> pointMap;
998 for(
int iband=0;iband<testImage.nrOfBand();++iband){
1000 testImage.readData(value,GDT_Float64,static_cast<int>(activePoints[iactive].posx),static_cast<int>(activePoints[iactive].posy),iband);
1003 pointMap[fs.str()]=value;
1005 pointMap[label_opt[0]]=0;
1007 testImage.image2geo(activePoints[iactive].posx,activePoints[iactive].posy,x,y);
1008 std::string fieldname=
"id";
1009 activeWriter.addPoint(x,y,pointMap,fieldname,++nactive);
1016 if(priorimg_opt.size())
1017 priorReader.close();
1020 if(entropy_opt.size())
1021 entropyImage.close();
1022 if(classBag_opt.size())
1023 classImageBag.close();
1024 classImageOut.close();
1029 for(
int ivalidation=0;ivalidation<input_opt.size();++ivalidation){
1030 if(output_opt.size())
1031 assert(output_opt.size()==input_opt.size());
1033 std::cout <<
"opening img reader " << input_opt[ivalidation] << std::endl;
1034 imgReaderOgr.open(input_opt[ivalidation]);
1037 if(output_opt.size()){
1039 std::cout <<
"opening img writer and copying fields from img reader" << output_opt[ivalidation] << std::endl;
1040 imgWriterOgr.open(output_opt[ivalidation],imgReaderOgr);
1043 cout <<
"number of layers in input ogr file: " << imgReaderOgr.getLayerCount() << endl;
1044 for(
int ilayer=0;ilayer<imgReaderOgr.getLayerCount();++ilayer){
1046 cout <<
"processing input layer " << ilayer << endl;
1047 if(output_opt.size()){
1049 std::cout <<
"creating field class" << std::endl;
1050 if(classValueMap.size())
1051 imgWriterOgr.createField(
"class",OFTInteger,ilayer);
1053 imgWriterOgr.createField(
"class",OFTString,ilayer);
1055 unsigned int nFeatures=imgReaderOgr.getFeatureCount(ilayer);
1056 unsigned int ifeature=0;
1058 pfnProgress(progress,pszMessage,pProgressArg);
1059 OGRFeature *poFeature;
1060 while( (poFeature = imgReaderOgr.getLayer(ilayer)->GetNextFeature()) != NULL ){
1061 if(verbose_opt[0]>1)
1062 std::cout <<
"feature " << ifeature << std::endl;
1063 if( poFeature == NULL ){
1064 cout <<
"Warning: could not read feature " << ifeature <<
" in layer " << imgReaderOgr.getLayerName(ilayer) << endl;
1067 OGRFeature *poDstFeature = NULL;
1068 if(output_opt.size()){
1069 poDstFeature=imgWriterOgr.createFeature(ilayer);
1070 if( poDstFeature->SetFrom( poFeature, TRUE ) != OGRERR_NONE ){
1071 CPLError( CE_Failure, CPLE_AppDefined,
1072 "Unable to translate feature %d from layer %s.\n",
1073 poFeature->GetFID(), imgWriterOgr.getLayerName(ilayer).c_str() );
1074 OGRFeature::DestroyFeature( poFeature );
1075 OGRFeature::DestroyFeature( poDstFeature );
1078 vector<float> validationPixel;
1079 vector<float> validationFeature;
1081 imgReaderOgr.readData(validationPixel,OFTReal,fields,poFeature,ilayer);
1082 assert(validationPixel.size()==nband);
1083 vector<float> probOut(nclass);
1084 for(
short iclass=0;iclass<nclass;++iclass)
1086 for(
int ibag=0;ibag<nbag;++ibag){
1087 for(
int iband=0;iband<nband;++iband){
1088 validationFeature.push_back((validationPixel[iband]-offset[ibag][iband])/scale[ibag][iband]);
1089 if(verbose_opt[0]==2)
1090 std::cout <<
" " << validationFeature.back();
1092 if(verbose_opt[0]==2)
1093 std::cout << std::endl;
1094 vector<double> result(nclass);
1096 x = (
struct svm_node *) malloc((validationFeature.size()+1)*
sizeof(
struct svm_node));
1097 for(
int i=0;i<validationFeature.size();++i){
1099 x[i].value=validationFeature[i];
1102 x[validationFeature.size()].index=-1;
1103 double predict_label=0;
1104 if(!prob_est_opt[0]){
1105 predict_label = svm_predict(svm[ibag],x);
1106 for(
short iclass=0;iclass<nclass;++iclass){
1107 if(iclass==static_cast<short>(predict_label))
1114 assert(svm_check_probability_model(svm[ibag]));
1115 predict_label = svm_predict_probability(svm[ibag],x,&(result[0]));
1117 if(verbose_opt[0]>1){
1118 std::cout <<
"predict_label: " << predict_label << std::endl;
1119 for(
int iclass=0;iclass<result.size();++iclass)
1120 std::cout << result[iclass] <<
" ";
1121 std::cout << std::endl;
1125 for(
short iclass=0;iclass<nclass;++iclass){
1126 switch(comb_opt[0]){
1129 probOut[iclass]+=result[iclass]*priors[iclass];
1132 probOut[iclass]*=pow(static_cast<float>(priors[iclass]),static_cast<float>(1.0-nbag)/nbag)*result[iclass];
1135 if(priors[iclass]*result[iclass]>probOut[iclass])
1136 probOut[iclass]=priors[iclass]*result[iclass];
1146 string classOut=
"Unclassified";
1147 for(
short iclass=0;iclass<nclass;++iclass){
1148 if(verbose_opt[0]>1)
1149 std::cout << probOut[iclass] <<
" ";
1150 if(probOut[iclass]>maxBag){
1151 maxBag=probOut[iclass];
1152 classOut=nameVector[iclass];
1156 if(verbose_opt[0]>1){
1157 if(classValueMap.size())
1158 std::cout <<
"->" << classValueMap[classOut] << std::endl;
1160 std::cout <<
"->" << classOut << std::endl;
1162 if(output_opt.size()){
1163 if(classValueMap.size())
1164 poDstFeature->SetField(
"class",classValueMap[classOut]);
1166 poDstFeature->SetField(
"class",classOut.c_str());
1167 poDstFeature->SetFID( poFeature->GetFID() );
1169 int labelIndex=poFeature->GetFieldIndex(label_opt[0].c_str());
1171 string classRef=poFeature->GetFieldAsString(labelIndex);
1173 if(classValueMap.size())
1174 cm.incrementResult(type2string<short>(classValueMap[classRef]),type2string<short>(classValueMap[classOut]),1);
1176 cm.incrementResult(classRef,classOut,1);
1180 if(output_opt.size()){
1181 if(imgWriterOgr.createFeature(poDstFeature,ilayer) != OGRERR_NONE){
1182 CPLError( CE_Failure, CPLE_AppDefined,
1183 "Unable to translate feature %d from layer %s.\n",
1184 poFeature->GetFID(), imgWriterOgr.getLayerName(ilayer).c_str() );
1185 OGRFeature::DestroyFeature( poDstFeature );
1186 OGRFeature::DestroyFeature( poDstFeature );
1190 if(!verbose_opt[0]){
1191 progress=
static_cast<float>(ifeature+1.0)/nFeatures;
1192 pfnProgress(progress,pszMessage,pProgressArg);
1194 OGRFeature::DestroyFeature( poFeature );
1195 OGRFeature::DestroyFeature( poDstFeature );
1198 imgReaderOgr.close();
1199 if(output_opt.size())
1200 imgWriterOgr.close();
1202 if(cm.nReference()){
1203 std::cout << cm << std::endl;
1204 cout <<
"class #samples userAcc prodAcc" << endl;
1211 for(
short iclass=0;iclass<cm.nClasses();++iclass){
1212 dua=cm.ua_pct(cm.getClass(iclass),&se95_ua);
1213 dpa=cm.pa_pct(cm.getClass(iclass),&se95_pa);
1214 cout << cm.getClass(iclass) <<
" " << cm.nReference(cm.getClass(iclass)) <<
" " << dua <<
" (" << se95_ua <<
")" <<
" " << dpa <<
" (" << se95_pa <<
")" << endl;
1216 std::cout <<
"Kappa: " << cm.kappa() << std::endl;
1217 doa=cm.oa(&se95_oa);
1218 std::cout <<
"Overall Accuracy: " << 100*doa <<
" (" << 100*se95_oa <<
")" << std::endl;
1222 if(active_opt.size())
1223 activeWriter.close();
1225 catch(
string errorString){
1226 std::cerr <<
"Error: errorString" << std::endl;
1229 for(
int ibag=0;ibag<nbag;++ibag){
1231 svm_destroy_param(¶m[ibag]);
1234 free(x_space[ibag]);
1235 svm_free_and_destroy_model(&(svm[ibag]));
throw this class when syntax error in command line option