otbListSampleGenerator.hxx

Go to the documentation of this file.

/*
 * Copyright (C) 2005-2024 Centre National d'Etudes Spatiales (CNES)
 *
 * This file is part of Orfeo Toolbox
 *
 *     https://www.orfeo-toolbox.org/
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
 
#ifndef otbListSampleGenerator_hxx
#define otbListSampleGenerator_hxx
 
#include "otbListSampleGenerator.h"
 
#include "itkImageRegionConstIteratorWithIndex.h"
#include "otbVectorDataProjectionFilter.h"
 
#include "otbMacro.h"
 
namespace otb
{
 
template <class TImage, class TVectorData>
ListSampleGenerator<TImage, TVectorData>::ListSampleGenerator()
  : m_MaxTrainingSize(-1),
    m_MaxValidationSize(-1),
    m_ValidationTrainingProportion(0.0),
    m_BoundByMin(true),
    m_PolygonEdgeInclusion(false),
    m_NumberOfClasses(0),
    m_ClassKey("Class"),
    m_ClassMinSize(-1)
{
  this->SetNumberOfRequiredInputs(2);
  this->SetNumberOfRequiredOutputs(4);
 
  // Register the outputs
  this->itk::ProcessObject::SetNthOutput(0, this->MakeOutput(0).GetPointer());
  this->itk::ProcessObject::SetNthOutput(1, this->MakeOutput(1).GetPointer());
  this->itk::ProcessObject::SetNthOutput(2, this->MakeOutput(2).GetPointer());
  this->itk::ProcessObject::SetNthOutput(3, this->MakeOutput(3).GetPointer());
 
  m_RandomGenerator = RandomGeneratorType::GetInstance();
}
 
template <class TImage, class TVectorData>
void ListSampleGenerator<TImage, TVectorData>::SetInput(const ImageType* image)
{
  this->ProcessObject::SetNthInput(0, const_cast<ImageType*>(image));
}
 
template <class TImage, class TVectorData>
const TImage* ListSampleGenerator<TImage, TVectorData>::GetInput() const
{
  if (this->GetNumberOfInputs() < 1)
  {
    return nullptr;
  }
 
  return static_cast<const ImageType*>(this->ProcessObject::GetInput(0));
}
 
template <class TImage, class TVectorData>
void ListSampleGenerator<TImage, TVectorData>::SetInputVectorData(const VectorDataType* vectorData)
{
  this->ProcessObject::SetNthInput(1, const_cast<VectorDataType*>(vectorData));
 
  // printVectorData(vectorData);
}
 
template <class TImage, class TVectorData>
const TVectorData* ListSampleGenerator<TImage, TVectorData>::GetInputVectorData() const
{
  if (this->GetNumberOfInputs() < 2)
  {
    return nullptr;
  }
 
  return static_cast<const VectorDataType*>(this->ProcessObject::GetInput(1));
}
 
template <class TImage, class TVectorData>
typename ListSampleGenerator<TImage, TVectorData>::DataObjectPointer ListSampleGenerator<TImage, TVectorData>::MakeOutput(DataObjectPointerArraySizeType idx)
{
  DataObjectPointer output;
  switch (idx)
  {
  case 0:
    output = static_cast<itk::DataObject*>(ListSampleType::New().GetPointer());
    break;
  case 1:
    output = static_cast<itk::DataObject*>(ListLabelType::New().GetPointer());
    break;
  case 2:
    output = static_cast<itk::DataObject*>(ListSampleType::New().GetPointer());
    break;
  case 3:
    output = static_cast<itk::DataObject*>(ListLabelType::New().GetPointer());
    break;
  default:
    output = static_cast<itk::DataObject*>(ListSampleType::New().GetPointer());
    break;
  }
  return output;
}
// Get the Training ListSample
template <class TImage, class TVectorData>
typename ListSampleGenerator<TImage, TVectorData>::ListSampleType* ListSampleGenerator<TImage, TVectorData>::GetTrainingListSample()
{
  return dynamic_cast<ListSampleType*>(this->itk::ProcessObject::GetOutput(0));
}
// Get the Training label ListSample
template <class TImage, class TVectorData>
typename ListSampleGenerator<TImage, TVectorData>::ListLabelType* ListSampleGenerator<TImage, TVectorData>::GetTrainingListLabel()
{
  return dynamic_cast<ListLabelType*>(this->itk::ProcessObject::GetOutput(1));
}
 
// Get the validation ListSample
template <class TImage, class TVectorData>
typename ListSampleGenerator<TImage, TVectorData>::ListSampleType* ListSampleGenerator<TImage, TVectorData>::GetValidationListSample()
{
  return dynamic_cast<ListSampleType*>(this->itk::ProcessObject::GetOutput(2));
}
 
 
// Get the validation label ListSample
template <class TImage, class TVectorData>
typename ListSampleGenerator<TImage, TVectorData>::ListLabelType* ListSampleGenerator<TImage, TVectorData>::GetValidationListLabel()
{
  return dynamic_cast<ListLabelType*>(this->itk::ProcessObject::GetOutput(3));
}
 
template <class TImage, class TVectorData>
void ListSampleGenerator<TImage, TVectorData>::GenerateInputRequestedRegion()
{
  ImagePointerType img = static_cast<ImageType*>(this->ProcessObject::GetInput(0));
 
  if (img.IsNotNull())
  {
 
    // Requested regions will be generated during GenerateData
    // call. For now request an empty region so as to avoid requesting
    // the largest possible region (fixes bug #943 )
    typename ImageType::RegionType dummyRegion;
    typename ImageType::SizeType   dummySize;
    dummySize.Fill(0);
    dummyRegion.SetSize(dummySize);
    img->SetRequestedRegion(dummyRegion);
  }
}
 
 
template <class TImage, class TVectorData>
void ListSampleGenerator<TImage, TVectorData>::GenerateData()
{
  // Get the inputs
  ImagePointerType      image      = const_cast<ImageType*>(this->GetInput());
  VectorDataPointerType vectorData = const_cast<VectorDataType*>(this->GetInputVectorData());
 
  // Get the outputs
  ListSamplePointerType trainingListSample   = this->GetTrainingListSample();
  ListLabelPointerType  trainingListLabel    = this->GetTrainingListLabel();
  ListSamplePointerType validationListSample = this->GetValidationListSample();
  ListLabelPointerType  validationListLabel  = this->GetValidationListLabel();
 
  // Gather some information about the relative size of the classes
  // We would like to have the same number of samples per class
  this->GenerateClassStatistics();
 
  this->ComputeClassSelectionProbability();
 
  // Clear the sample lists
  trainingListSample->Clear();
  trainingListLabel->Clear();
  validationListSample->Clear();
  validationListLabel->Clear();
 
  // Set MeasurementVectorSize for each sample list
  trainingListSample->SetMeasurementVectorSize(image->GetNumberOfComponentsPerPixel());
  // stores label as integers,so put the size to 1
  trainingListLabel->SetMeasurementVectorSize(1);
  validationListSample->SetMeasurementVectorSize(image->GetNumberOfComponentsPerPixel());
  // stores label as integers,so put the size to 1
  validationListLabel->SetMeasurementVectorSize(1);
 
  m_ClassesSamplesNumberTraining.clear();
  m_ClassesSamplesNumberValidation.clear();
 
  typename ImageType::RegionType imageLargestRegion = image->GetLargestPossibleRegion();
 
  auto itVectorPair = vectorData->GetIteratorPair();
  auto currentIt = itVectorPair.first;
  for (; currentIt != itVectorPair.second; ++currentIt)
  {
    if (vectorData->Get(currentIt)->IsPolygonFeature())
    {
      PolygonPointerType exteriorRing = vectorData->Get(currentIt)->GetPolygonExteriorRing();
 
      typename ImageType::RegionType polygonRegion = otb::TransformPhysicalRegionToIndexRegion(exteriorRing->GetBoundingRegion(), image.GetPointer());
 
      const bool hasIntersection = polygonRegion.Crop(imageLargestRegion);
      if (!hasIntersection)
      {
        continue;
      }
 
      image->SetRequestedRegion(polygonRegion);
      image->PropagateRequestedRegion();
      image->UpdateOutputData();
 
      typedef itk::ImageRegionConstIteratorWithIndex<ImageType> IteratorType;
      IteratorType                                              it(image, polygonRegion);
 
      for (it.GoToBegin(); !it.IsAtEnd(); ++it)
      {
        itk::ContinuousIndex<double, 2> point;
        image->TransformIndexToPhysicalPoint(it.GetIndex(), point);
 
        if (exteriorRing->IsInside(point) || (this->GetPolygonEdgeInclusion() && exteriorRing->IsOnEdge(point)))
        {
          PolygonListPointerType interiorRings = vectorData->Get(currentIt)->GetPolygonInteriorRings();
 
          bool isInsideInteriorRing = false;
          for (typename PolygonListType::Iterator interiorRing = interiorRings->Begin(); interiorRing != interiorRings->End(); ++interiorRing)
          {
            if (interiorRing.Get()->IsInside(point) || (this->GetPolygonEdgeInclusion() && interiorRing.Get()->IsOnEdge(point)))
            {
              isInsideInteriorRing = true;
              break;
            }
          }
          if (isInsideInteriorRing)
          {
            continue; // skip this pixel and continue
          }
 
          double randomValue = m_RandomGenerator->GetUniformVariate(0.0, 1.0);
          if (randomValue < m_ClassesProbTraining[vectorData->Get(currentIt)->GetFieldAsInt(m_ClassKey)])
          {
            // Add the sample to the training list
            trainingListSample->PushBack(it.Get());
            trainingListLabel->PushBack(vectorData->Get(currentIt)->GetFieldAsInt(m_ClassKey));
            m_ClassesSamplesNumberTraining[vectorData->Get(currentIt)->GetFieldAsInt(m_ClassKey)] += 1;
          }
          else if (randomValue <
                   m_ClassesProbTraining[vectorData->Get(currentIt)->GetFieldAsInt(m_ClassKey)] + m_ClassesProbValidation[vectorData->Get(currentIt)->GetFieldAsInt(m_ClassKey)])
          {
            // Add the sample to the validation list
            validationListSample->PushBack(it.Get());
            validationListLabel->PushBack(vectorData->Get(currentIt)->GetFieldAsInt(m_ClassKey));
            m_ClassesSamplesNumberValidation[vectorData->Get(currentIt)->GetFieldAsInt(m_ClassKey)] += 1;
          }
          // Note: some samples may not be used at all
        }
      }
    }
  }
 
  assert(trainingListSample->Size() == trainingListLabel->Size());
  assert(validationListSample->Size() == validationListLabel->Size());
  this->UpdateProgress(1.0f);
}
 
template <class TImage, class TVectorData>
void ListSampleGenerator<TImage, TVectorData>::GenerateClassStatistics()
{
  m_ClassesSize.clear();
 
  ImageType*                            image      = const_cast<ImageType*>(this->GetInput());
  typename VectorDataType::ConstPointer vectorData = this->GetInputVectorData();
 
  // Compute cumulative area of all polygons of each class
  auto itVectorPair = vectorData->GetIteratorPair();
  auto currentIt = itVectorPair.first;
  for (; currentIt != itVectorPair.second; ++currentIt)
  {
    typename DataNodeType::Pointer datanode = vectorData->Get(currentIt);
    if (datanode->IsPolygonFeature())
    {
      double area = GetPolygonAreaInPixelsUnits(datanode, image);
      m_ClassesSize[datanode->GetFieldAsInt(m_ClassKey)] += area;
    }
  }
  m_NumberOfClasses = m_ClassesSize.size();
}
 
template <class TImage, class TVectorData>
void ListSampleGenerator<TImage, TVectorData>::ComputeClassSelectionProbability()
{
  m_ClassesProbTraining.clear();
  m_ClassesProbValidation.clear();
 
  // Sanity check
  if (m_ClassesSize.empty())
  {
    itkGenericExceptionMacro(<< "No training sample found inside image");
  }
 
  // Go through the classes size to find the smallest one
  double minSize = itk::NumericTraits<double>::max();
  for (std::map<ClassLabelType, double>::const_iterator itmap = m_ClassesSize.begin(); itmap != m_ClassesSize.end(); ++itmap)
  {
    if (minSize > itmap->second)
    {
      minSize = itmap->second;
    }
  }
 
  // Apply the proportion between training and validation samples (all training by default)
  double minSizeTraining   = minSize * (1.0 - m_ValidationTrainingProportion);
  double minSizeValidation = minSize * m_ValidationTrainingProportion;
 
  // Apply the limit if specified by the user
  if (m_BoundByMin)
  {
    if ((m_MaxTrainingSize != -1) && (m_MaxTrainingSize < minSizeTraining))
    {
      minSizeTraining = m_MaxTrainingSize;
    }
    if ((m_MaxValidationSize != -1) && (m_MaxValidationSize < minSizeValidation))
    {
      minSizeValidation = m_MaxValidationSize;
    }
  }
  // Compute the probability selection for each class
  for (std::map<ClassLabelType, double>::const_iterator itmap = m_ClassesSize.begin(); itmap != m_ClassesSize.end(); ++itmap)
  {
    if (m_BoundByMin)
    {
      m_ClassesProbTraining[itmap->first]   = minSizeTraining / itmap->second;
      m_ClassesProbValidation[itmap->first] = minSizeValidation / itmap->second;
    }
    else
    {
      long int maxSizeT = (itmap->second) * (1.0 - m_ValidationTrainingProportion);
      long int maxSizeV = (itmap->second) * m_ValidationTrainingProportion;
 
      // Check if max sizes respect the maximum bounds
      double correctionRatioTrain = 1.0;
      if ((m_MaxTrainingSize > -1) && (m_MaxTrainingSize < maxSizeT))
      {
        correctionRatioTrain = (double)(m_MaxTrainingSize) / (double)(maxSizeT);
      }
      double correctionRatioValid = 1.0;
      if ((m_MaxValidationSize > -1) && (m_MaxValidationSize < maxSizeV))
      {
        correctionRatioValid = (double)(m_MaxValidationSize) / (double)(maxSizeV);
      }
      double correctionRatio                = std::min(correctionRatioTrain, correctionRatioValid);
      m_ClassesProbTraining[itmap->first]   = correctionRatio * (1.0 - m_ValidationTrainingProportion);
      m_ClassesProbValidation[itmap->first] = correctionRatio * m_ValidationTrainingProportion;
    }
  }
}
template <class TImage, class TVectorData>
double ListSampleGenerator<TImage, TVectorData>::GetPolygonAreaInPixelsUnits(DataNodeType* polygonDataNode, ImageType* image)
{
  const double pixelArea = std::abs(image->GetSignedSpacing()[0] * image->GetSignedSpacing()[1]);
 
  // Compute area of exterior ring in pixels
  PolygonPointerType exteriorRing = polygonDataNode->GetPolygonExteriorRing();
  double             area         = exteriorRing->GetArea() / pixelArea;
 
  // Remove contribution of all interior rings
  PolygonListPointerType interiorRings = polygonDataNode->GetPolygonInteriorRings();
  for (typename PolygonListType::Iterator interiorRing = interiorRings->Begin(); interiorRing != interiorRings->End(); ++interiorRing)
  {
    area -= interiorRing.Get()->GetArea() / pixelArea;
  }
 
  return area;
}
 
template <class TImage, class TVectorData>
void ListSampleGenerator<TImage, TVectorData>::PrintSelf(std::ostream& os, itk::Indent indent) const
{
  os << indent << "* MaxTrainingSize: " << m_MaxTrainingSize << "\n";
  os << indent << "* MaxValidationSize: " << m_MaxValidationSize << "\n";
  os << indent << "* Proportion: " << m_ValidationTrainingProportion << "\n";
  os << indent << "* Input data:\n";
  if (m_ClassesSize.empty())
  {
    os << indent << "Empty\n";
  }
  else
  {
    for (std::map<ClassLabelType, double>::const_iterator itmap = m_ClassesSize.begin(); itmap != m_ClassesSize.end(); ++itmap)
    {
      os << indent << itmap->first << ": " << itmap->second << "\n";
    }
  }
 
  os << "\n" << indent << "* Training set:\n";
  if (m_ClassesProbTraining.empty())
  {
    os << indent << "Not computed\n";
  }
  else
  {
    os << indent << "** Selection probability:\n";
    for (std::map<ClassLabelType, double>::const_iterator itmap = m_ClassesProbTraining.begin(); itmap != m_ClassesProbTraining.end(); ++itmap)
    {
      os << indent << itmap->first << ": " << itmap->second << "\n";
    }
    os << indent << "** Number of selected samples:\n";
    for (std::map<ClassLabelType, int>::const_iterator itmap = m_ClassesSamplesNumberTraining.begin(); itmap != m_ClassesSamplesNumberTraining.end(); ++itmap)
    {
      os << indent << itmap->first << ": " << itmap->second << "\n";
    }
  }
 
  os << "\n" << indent << "* Validation set:\n";
  if (m_ClassesProbValidation.empty())
  {
    os << indent << "Not computed\n";
  }
  else
  {
    os << indent << "** Selection probability:\n";
    for (std::map<ClassLabelType, double>::const_iterator itmap = m_ClassesProbValidation.begin(); itmap != m_ClassesProbValidation.end(); ++itmap)
    {
      os << indent << itmap->first << ": " << itmap->second << "\n";
    }
    os << indent << "** Number of selected samples:\n";
    for (std::map<ClassLabelType, int>::const_iterator itmap = m_ClassesSamplesNumberValidation.begin(); itmap != m_ClassesSamplesNumberValidation.end();
         ++itmap)
    {
      os << indent << itmap->first << ": " << itmap->second << "\n";
    }
  }
}
}
 
#endif