// Multiply by vector
NEWMAT::ReturnMatrix FullBFMatrix::MulByVec(const NEWMAT::ColumnVector& invec) const
{
if (invec.Nrows() != int(Ncols())) {throw BFMatrixException("FullBFMatrix::MulByVec: Matrix-vector size mismatch");}
NEWMAT::ColumnVector ret;
ret = (*mp)*invec;
ret.Release();
return(ret);
}
// Given A*x=b, solve for x
NEWMAT::ReturnMatrix FullBFMatrix::SolveForx(const NEWMAT::ColumnVector& b, // Ignoring all parameters except b
MISCMATHS::MatrixType type,
double tol,
int miter) const
{
if (int(Nrows()) != b.Nrows()) {throw BFMatrixException("FullBFMatrix::SolveForx: Matrix-vector size mismatch");}
NEWMAT::ColumnVector ret;
ret = mp->i()*b;
ret.Release();
return(ret);
}
NEWMAT::ColumnVector
ossimRpcProjection::getResidue(const ossimTieGptSet& tieSet)
{
//init
NEWMAT::ColumnVector residue;
int dimObs;
bool useImageObs = useForward(); //caching
if (useImageObs)
{
dimObs = 2; //image observation
} else {
dimObs = 3; //ground observations
}
int no = dimObs * tieSet.size(); //number of observations
residue.ReSize(no);
const vector<ossimRefPtr<ossimTieGpt> >& theTPV = tieSet.getTiePoints();
vector<ossimRefPtr<ossimTieGpt> >::const_iterator tit;
unsigned long c=1;
if (useImageObs)
{
//image observations
ossimDpt resIm;
// loop on tie points
for (tit = theTPV.begin() ; tit != theTPV.end() ; ++tit)
{
//compute residue
resIm = (*tit)->tie - forward(**tit);
residue(c++) = resIm.x;
residue(c++) = resIm.y;
}
} else {
// ground observations
ossimGpt gd;
// loop on tie points
for (tit = theTPV.begin() ; tit != theTPV.end() ; ++tit)
{
//compute residue
gd = inverse((*tit)->tie);
residue(c++) = ((*tit)->lon - gd.lon) * 100000.0; //approx meters //TBC TBD
residue(c++) = ((*tit)->lat - gd.lat) * 100000.0 * cos(gd.lat / 180.0 * M_PI);
residue(c++) = (*tit)->hgt - gd.hgt; //meters
}
} //end of if (useImageObs)
return residue;
}
void energyEvalFunction_init(int ndim, NEWMAT::ColumnVector& x)
{
for (int i = 0; i < ndim; i++)
{
x.element(i) = opt_params.orig_params_each_piece->element(i);
}
}
void threadParamEvalFunction_init(int ndim, NEWMAT::ColumnVector& x)
{
for (int i = 0; i < ndim; i++)
{
x.element(i) = opt_params_thread_params.orig_params->element(i);
}
}
phy::vector_t const toNumber(NEWMAT::ColumnVector const & v)
{
unsigned n = v.Nrows();
phy::vector_t u(n);
for (unsigned i = 0; i < n; i++)
u[i] = v[i];
return u;
}
void
nlf1(int mode, int ndim, NEWMAT::ColumnVector const & x, OPTPP::real & fx, NEWMAT::ColumnVector & gx, int & result, void * data)
{
result = OPTPP::NLPNoOp;
if (mode & OPTPP::NLPFunction)
nlf0(ndim, x, fx, result, data);
if (mode & OPTPP::NLPGradient)
{
ScalarFunction const * func = static_cast<ScalarFunction *>(data);
AnalyticD1ScalarFunction const * func1 = dynamic_cast<AnalyticD1ScalarFunction const *>(func);
alwaysAssertM(func1, "OPTPPNumericalOptimizer: Function does not have analytical first derivative");
gx.ReSize(ndim);
func1->gradientAt(x.data(), gx.data());
result |= OPTPP::NLPGradient;
}
}
void CUniversalKriging::Init_r(const CGridPoint& pt, CGridPointVector& va, NEWMAT::ColumnVector& r)const
{
double unbias = m_pVariogram->GetUnbias();
size_t neq = va.size() + GetMDT();
r.resize((int)neq);
r = 0;
CDiscretisation m_discr;
m_discr.Initialization(pt, *m_pVariogram, m_p, m_param.m_bFillNugget);
// Set up the right hand side:
for (int i = 0; i < (int)va.size(); i++)
{
//compute mean over dicretisation
double cb = 0.0;
for (int j = 0; j < (int)m_discr.size(); j++)
{
cb += m_pVariogram->cova3(va[i], m_discr[j]);
double d = va[i].GetDistance(va[j]);
if (d < EPSLON && (m_discr.size() > 1 || m_param.m_bFillNugget))
cb -= m_pVariogram->GetNugget();
}
cb = cb / m_discr.size();
r[i] = cb;
}
// Fill in the OK unbiasedness portion of the matrix (if not doing SK):
if (!m_p.m_bSK)
r[(int)va.size()] = unbias;
// Add the additional Unbiasedness constraints:
int im = int(va.size()) + 1;
double resc = max(m_pVariogram->GetMaxCov(), LIMIT) / (2.0 * m_p.m_radius);
const CDetrending& detrending = m_pVariogram->GetDetrending();
for (int i = 0; i < m_externalDrift.size(); i++)
{
r[im] = m_pVariogram->GetMaxCov();
im++;
}
}
void
nlf2(int mode, int ndim, NEWMAT::ColumnVector const & x, OPTPP::real & fx, NEWMAT::ColumnVector & gx,
NEWMAT::SymmetricMatrix & Hx, int & result, void * data)
{
result = OPTPP::NLPNoOp;
if (mode & OPTPP::NLPFunction || mode & OPTPP::NLPGradient)
nlf1(mode, ndim, x, fx, gx, result, data);
if (mode & OPTPP::NLPHessian)
{
ScalarFunction const * func = static_cast<ScalarFunction *>(data);
AnalyticD2ScalarFunction const * func2 = dynamic_cast<AnalyticD2ScalarFunction const *>(func);
alwaysAssertM(func2, "OPTPPNumericalOptimizer: Function does not have analytical second derivative");
Hx.ReSize(ndim);
OPTPPSymMatWrapper Hwrapper(&Hx);
func2->hessianAt(x.data(), Hwrapper);
result |= OPTPP::NLPHessian;
}
}
void
ossimRpcProjection::buildNormalEquation(const ossimTieGptSet& tieSet,
NEWMAT::SymmetricMatrix& A,
NEWMAT::ColumnVector& residue,
NEWMAT::ColumnVector& projResidue,
double pstep_scale)
{
//goal: build Least Squares system
//constraint: never store full Jacobian matrix in memory (can be huge)
// so we build the matrices incrementally
// the system can be built using forward() or inverse() depending on the projection capabilities : useForward()
//
//TBD : add covariance matrix for each tie point
//init
int np = getNumberOfAdjustableParameters();
int dimObs;
bool useImageObs = useForward(); //caching
if (useImageObs)
{
dimObs = 2; //image observation
} else {
dimObs = 3; //ground observations
}
int no = dimObs * tieSet.size(); //number of observations
A.ReSize(np);
residue.ReSize(no);
projResidue.ReSize(np);
//Zeroify matrices that will be accumulated
A = 0.0;
projResidue = 0.0;
const vector<ossimRefPtr<ossimTieGpt> >& theTPV = tieSet.getTiePoints();
vector<ossimRefPtr<ossimTieGpt> >::const_iterator tit;
unsigned long c=1;
if (useImageObs)
{
//image observations
ossimDpt* imDerp = new ossimDpt[np];
ossimDpt resIm;
// loop on tie points
for (tit = theTPV.begin() ; tit != theTPV.end() ; ++tit)
{
//compute residue
resIm = (*tit)->tie - forward(*(*tit));
residue(c++) = resIm.x;
residue(c++) = resIm.y;
//compute all image derivatives regarding parametres for the tie point position
for(int p=0;p<np;++p)
{
imDerp[p] = getForwardDeriv( p , *(*tit) , pstep_scale);
}
//compute influence of tie point on all sytem elements
for(int p1=0;p1<np;++p1)
{
//proj residue: J * residue
projResidue.element(p1) += imDerp[p1].x * resIm.x + imDerp[p1].y * resIm.y;
//normal matrix A = transpose(J)*J
for(int p2=p1;p2<np;++p2)
{
A.element(p1,p2) += imDerp[p1].x * imDerp[p2].x + imDerp[p1].y * imDerp[p2].y;
}
}
}
delete []imDerp;
}
else
{
// ground observations
std::vector<ossimGpt> gdDerp(np);
ossimGpt gd, resGd;
// loop on tie points
for (tit = theTPV.begin() ; tit != theTPV.end() ; ++tit)
{
//compute residue
gd = inverse((*tit)->tie);
residue(c++) = resGd.lon = ((*tit)->lon - gd.lon) * 100000.0;
residue(c++) = resGd.lat = ((*tit)->lat - gd.lat) * 100000.0 * cos(gd.lat / 180.0 * M_PI);
residue(c++) = resGd.hgt = (*tit)->hgt - gd.hgt; //TBD : normalize to meters?
//compute all image derivatives regarding parametres for the tie point position
for(int p=0;p<np;++p)
{
gdDerp[p] = getInverseDeriv( p , (*tit)->tie, pstep_scale);
}
//compute influence of tie point on all sytem elements
for(int p1=0;p1<np;++p1)
{
//proj residue: J * residue
projResidue.element(p1) += gdDerp[p1].lon * resGd.lon + gdDerp[p1].lat * resGd.lat + gdDerp[p1].hgt * resGd.hgt; //TBC
//normal matrix A = transpose(J)*J
for(int p2=p1;p2<np;++p2)
{
A.element(p1,p2) += gdDerp[p1].lon * gdDerp[p2].lon + gdDerp[p1].lat * gdDerp[p2].lat + gdDerp[p1].hgt * gdDerp[p2].hgt;
}
}
}
} //end of if (useImageObs)
}
double
ossimRpcProjection::optimizeFit(const ossimTieGptSet& tieSet, double* /* targetVariance */)
{
#if 1
//NOTE : ignore targetVariance
ossimRefPtr<ossimRpcSolver> solver = new ossimRpcSolver(false, false); //TBD : choices should be part of setupFromString
std::vector<ossimDpt> imagePoints;
std::vector<ossimGpt> groundPoints;
tieSet.getSlaveMasterPoints(imagePoints, groundPoints);
solver->solveCoefficients(imagePoints, groundPoints);
ossimRefPtr< ossimImageGeometry > optProj = solver->createRpcProjection();
if (!optProj)
{
ossimNotify(ossimNotifyLevel_FATAL) << "FATAL ossimRpcProjection::optimizeFit(): error when optimizing the RPC with given tie points"
<< std::endl;
return -1.0;
}
if(optProj->hasProjection())
{
ossimKeywordlist kwl;
optProj->getProjection()->saveState(kwl);
this->loadState(kwl);
}
return std::pow(solver->getRmsError(), 2); //variance in pixel^2
#else
// COPIED from ossimRpcProjection
//
//
//use a simple Levenberg-Marquardt non-linear optimization
//note : please limit the number of tie points
//
//INPUTS: requires Jacobian matrix (partial derivatives with regards to parameters)
//OUTPUTS: will also compute parameter covariance matrix
//
//TBD: use targetVariance!
int np = getNumberOfAdjustableParameters();
int nobs = tieSet.size();
//setup initail values
int iter=0;
int iter_max = 200;
double minResidue = 1e-10; //TBC
double minDelta = 1e-10; //TBC
//build Least Squares initial normal equation
// don't waste memory, add samples one at a time
NEWMAT::SymmetricMatrix A;
NEWMAT::ColumnVector residue;
NEWMAT::ColumnVector projResidue;
double deltap_scale = 1e-4; //step_Scale is 1.0 because we expect parameters to be between -1 and 1
buildNormalEquation(tieSet, A, residue, projResidue, deltap_scale);
double ki2=residue.SumSquare();
//get current adjustment (between -1 and 1 normally) and convert to ColumnVector
ossimAdjustmentInfo cadj;
getAdjustment(cadj);
std::vector< ossimAdjustableParameterInfo >& parmlist = cadj.getParameterList();
NEWMAT::ColumnVector cparm(np), nparm(np);
for(int n=0;n<np;++n)
{
cparm(n+1) = parmlist[n].getParameter();
}
double damping_speed = 2.0;
//find max diag element for A
double maxdiag=0.0;
for(int d=1;d<=np;++d) {
if (maxdiag < A(d,d)) maxdiag=A(d,d);
}
double damping = 1e-3 * maxdiag;
double olddamping = 0.0;
bool found = false;
//DEBUG TBR
cout<<"rms="<<sqrt(ki2/nobs)<<" ";
cout.flush();
while ( (!found) && (iter < iter_max) ) //non linear optimization loop
{
bool decrease = false;
do
{
//add damping update to normal matrix
for(int d=1;d<=np;++d) A(d,d) += damping - olddamping;
olddamping = damping;
NEWMAT::ColumnVector deltap = solveLeastSquares(A, projResidue);
if (deltap.NormFrobenius() <= minDelta)
{
found = true;
} else {
//update adjustment
nparm = cparm + deltap;
for(int n=0;n<np;++n)
{
setAdjustableParameter(n, nparm(n+1), false); //do not update now, wait
}
adjustableParametersChanged();
//check residue is reduced
NEWMAT::ColumnVector newresidue = getResidue(tieSet);
double newki2=newresidue.SumSquare();
double res_reduction = (ki2 - newki2) / (deltap.t()*(deltap*damping + projResidue)).AsScalar();
//DEBUG TBR
cout<<sqrt(newki2/nobs)<<" ";
cout.flush();
if (res_reduction > 0)
{
//accept new parms
cparm = nparm;
ki2=newki2;
deltap_scale = max(1e-15, deltap.NormInfinity()*1e-4);
buildNormalEquation(tieSet, A, residue, projResidue, deltap_scale);
olddamping = 0.0;
found = ( projResidue.NormInfinity() <= minResidue );
//update damping factor
damping *= std::max( 1.0/3.0, 1.0-std::pow((2.0*res_reduction-1.0),3));
damping_speed = 2.0;
decrease = true;
} else {
//cancel parameter update
for(int n=0;n<np;++n)
{
setAdjustableParameter(n, nparm(n+1), false); //do not update right now
}
adjustableParametersChanged();
damping *= damping_speed;
damping_speed *= 2.0;
}
}
} while (!decrease && !found);
++iter;
}
//DEBUG TBR
cout<<endl;
//compute parameter correlation
// use normal matrix inverse
//TBD
return ki2/nobs;
#endif
}
void initWrapper(int ndim, NEWMAT::ColumnVector & x)
{
assert( ndim == x.Nrows() );
assert( ndim == static_cast<signed>( glbInitParams->size() ) );
x = toColumnVector( *glbInitParams );
}
double CUniversalKriging::Evaluate(const CGridPoint& pt, int iXval)const
{
//if variogram is not initialised correckly, we retuern no data;
if (!m_pVariogram->IsInit())
return m_param.m_noData;
// Find the nearest samples:
CGridPointVector va;
Init_va(pt, iXval, va);
int mdt = GetMDT();
// Test number of samples found:
// Test if there are enough samples to estimate all drift terms:
if ((int)va.size() < m_p.m_nbPoint || (int)va.size() <= mdt)
return m_param.m_noData;
if ( /*iXval<0 &&*/ va[0].GetDistance(pt) > m_param.m_maxDistance)
return m_param.m_noData;
// There are enough samples - proceed with estimation.
// Go ahead and set up the OK portion of the kriging matrix:
NEWMAT::ColumnVector r;
NEWMAT::Matrix a;
Init_a(pt, va, a);
Init_r(pt, va, r);
// If estimating the trend then reset all the right hand side terms=0.0:
if (m_p.m_bTrend)
r = 0;
//copy r to compute variance
NEWMAT::ColumnVector rr(r);
// Solve the kriging system:
int neq = mdt + int(va.size());
//reduce the matrix until the 2/3 of the number of neightbor
bool bOK = false;
while (!bOK && neq >= mdt + m_p.m_nbPoint * 2 / 3)//by RSA 25/10/2010
{
Try
{
a = a.i();
bOK = true;
}
Catch(NEWMAT::Exception)
{
OutputDebugStringW(L"a=a.i() failled; Reduce number of point");
Init_a(pt, va, a);
Init_r(pt, va, r);
neq--;
int firstRow = 1 + mdt + int(va.size()) - neq;
a = a.SubMatrix(firstRow, a.Nrows(), firstRow, a.Ncols());
r = r.SubMatrix(firstRow, r.Nrows(), 1, 1);
}
if (m_pAgent && m_pAgent->TestConnection(m_hxGridSessionID) == S_FALSE)
throw(CHxGridException());
}
//if we don't solve system, return missing
if (!bOK)
return m_param.m_noData;
NEWMAT::ColumnVector s = a*r;
CDiscretisation m_discr;
m_discr.Initialization(pt, *m_pVariogram, m_p, m_param.m_bFillNugget);
// Compute the solution:
double uuk = 0.0;
double ukv = m_discr.cbb;
for (int j = 0; j < neq; j++)
{
ukv -= s[j] * rr[j];
if (j < (int)va.size())
uuk += s[j] * (m_prePostTransfo.Transform(va[j].m_event) - m_p.m_SKmean);
}
uuk += m_p.m_SKmean;
uuk = m_prePostTransfo.InvertTransform(uuk, m_param.m_noData);
//
if ( /*iXval<0 &&*/ m_param.m_bRegionalLimit && uuk > m_param.m_noData)
{
CStatistic stat;
for (size_t i = 0; i < va.size(); i++)
stat += va[i].m_event;
bool bOutside = uuk<stat[LOWEST] - m_param.m_regionalLimitSD*stat[STD_DEV] || uuk>stat[HIGHEST] + m_param.m_regionalLimitSD*stat[STD_DEV];
if (bOutside)
{
if (m_param.m_bRegionalLimitToBound)
uuk = min(stat[HIGHEST] + m_param.m_regionalLimitSD*stat[STD_DEV], max(stat[LOWEST] - m_param.m_regionalLimitSD*stat[STD_DEV], uuk));
else
uuk = m_param.m_noData;
}
}
if ( /*iXval<0 &&*/ m_param.m_bGlobalLimit && uuk > m_param.m_noData)
{
bool bOutside = uuk<m_stat[LOWEST] - m_param.m_globalLimitSD*m_stat[STD_DEV] || uuk>m_stat[HIGHEST] + m_param.m_globalLimitSD*m_stat[STD_DEV];
if (bOutside)
{
if (m_param.m_bGlobalLimitToBound)
uuk = min(m_stat[HIGHEST] + m_param.m_globalLimitSD*m_stat[STD_DEV], max(m_stat[LOWEST] - m_param.m_globalLimitSD*m_stat[STD_DEV], uuk));
else
uuk = m_param.m_noData;
}
}
if ( /*iXval<0 &&*/ m_param.m_bGlobalMinMaxLimit && uuk > m_param.m_noData)
{
bool bOutside = uuk<m_param.m_globalMinLimit || uuk>m_param.m_globalMaxLimit;
if (bOutside)
{
if (m_param.m_bGlobalMinMaxLimitToBound)
uuk = min(m_param.m_globalMaxLimit, max(m_param.m_globalMinLimit, uuk));
else
uuk = m_param.m_noData;
}
}
return uuk;
}
