Una veloce presentazione Una veloce presentazione ( (Silvio GRIGUOLO - DAEST/IUAV ) ) Qualche slide iniziale è tratta da una presentazione di Qualche slide iniziale è tratta da una presentazione di Sander Mucher, Wageningen, coordinatore del progetto Sander Mucher, Wageningen, coordinatore del progetto Le ultime slides sull’applicazione alla biodiversità s Le ultime slides sull’applicazione alla biodiversità s di Camiel Heunks, del National Institute for public di Camiel Heunks, del National Institute for public Health and the Environment (RIVM) Health and the Environment (RIVM) La presentazione usa sia l’italiano che l’inglese La presentazione usa sia l’italiano che l’inglese
120
Embed
Una veloce presentazione () (Silvio GRIGUOLO - DAEST/IUAV) Qualche slide iniziale è tratta da una presentazione di Qualche slide iniziale è tratta da una.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Una veloce presentazioneUna veloce presentazione((Silvio GRIGUOLO - DAEST/IUAV))
• Qualche slide iniziale è tratta da una presentazione diQualche slide iniziale è tratta da una presentazione di
Sander Mucher, Wageningen, coordinatore del progettoSander Mucher, Wageningen, coordinatore del progetto
• Le ultime slides sull’applicazione alla biodiversità sonoLe ultime slides sull’applicazione alla biodiversità sono
di Camiel Heunks, del National Institute for publicdi Camiel Heunks, del National Institute for public
Health and the Environment (RIVM)Health and the Environment (RIVM)
• La presentazione usa sia l’italiano che l’ingleseLa presentazione usa sia l’italiano che l’inglese
Land Cover Classification at the Land Cover Classification at the continental scalecontinental scale
the PELCOM Projectthe PELCOM Project
A shared Cost Action, EU 4th framework Environment & Climate
Start September ‘96 - End November ‘99
• Wageningen UR ( formerly DLO)• Austrian Research Centre Seibersdorf (ARCS)• Centre National de Recherches Meteorologiques
(CNRM) - Toulouse• National Institute for Public Health and the
Environment (RIVM)• Swedish Space Corporation (SSC)• Instituto Universitario di Architettura (IUAV)• Space Applications Institute (SAI-JRC)• Geodan
PELCOMPELCOMPan European Land COver Monitoring(Research funded by the European Commission)
Perché il Progetto Perché il Progetto PELCOM?PELCOM?
Stato dell’arteStato dell’arte
• CORINE-LC procede a rilento• Altri database raster esistenti sono
largamente basati su statistiche amministrative
• La classificazione DISCover (IGBP) ha un livello di dettaglio insufficiente per l’Europa
• Il progetto FIRS si concentra soprattutto sulle foreste
CORINECORINE land cover database land cover database
Detail of IGBP global land cover Detail of IGBP global land cover databasedatabase
In definitiva...In definitiva...
Manca un’informazione aggiornata del Land Cover per l’intera Europa
Dunque…Dunque…investigare l’applicabilità delle immagini satellitari NOAA-AVHRR per:
• Mapping• Monitoring
Gli obiettivi del ProgettoGli obiettivi del Progetto• definizione di una metodologia di classificazione consistente
• Costruzione di una database raster delle coperture del suolo (Land Cover) ad 1 km dell’intera Europa, da utilizzare come input per modelli dinamici ambientali alla scala continentale
• Applicazioni sperimentali della mappa
Fasi (1996-99)Fasi (1996-99)
• 1st year sviluppo della metodologia di
classificazione e monitoraggio• 2nd year esperimenti regionali di
classificazione
• 3rd year Studio di casi su ambiente & clima
Primi esperimenti di Primi esperimenti di modellisticamodellistica
1 Modello dinamico della bio-diversità, basatoModello dinamico della bio-diversità, basato sulla definizione di sulla definizione di Natural Capital IndexNatural Capital Index (pressione sulla bio-diversità dovute a (pressione sulla bio-diversità dovute a cambiamenti climatici, densità di insediamentocambiamenti climatici, densità di insediamento umano, consumo e produzione, acidificazione,umano, consumo e produzione, acidificazione, eutrofizzazione, ozono, ecc.)eutrofizzazione, ozono, ecc.)
RIVM - Indice di RIVM - Indice di pressione sulla bio-pressione sulla bio-diversità nelle aree diversità nelle aree naturali naturali
(proiezione al 2010)(proiezione al 2010)
2 MeteoFrance (Toulouse)MeteoFrance (Toulouse):: applicazione a modelli metereologiciapplicazione a modelli metereologici.. Gli scambi vegetazione-atmosfera sono Gli scambi vegetazione-atmosfera sono unauna componente importante nei modelli componente importante nei modelli numerici dinumerici di previsione del tempo.previsione del tempo. La mappa PELCOM della vegetazione è La mappa PELCOM della vegetazione è statastata assunta come input per il modello assunta come input per il modello metereolo-metereolo- gico ARPEGE-ALADDIN. gico ARPEGE-ALADDIN.
VOC emission dalle foreste usando PELCOMVOC emission dalle foreste usando PELCOM (ARCS) (ARCS)
Le piante sono una delle maggiori fonti di Composti OrganiciVolatili, a loro volta precursori della formazione di ozono.L’emissione di composti organici è ritenuta una risposta fisiologica delle piante ai fattori di stree ambientale, comel’alta temperatura o la scarsità idrica.
I datiI dati
MaterialiMateriali• AVHRR NDVI
monthly MVC’s 1997
• AVHRR Daily Multi-spectral mosaics ‘95/’96/’97
• Ancillary geographic data:
• CORINE land cover data• FIRS regions and strata• Digital Chart of the World
Il Digital Terrain Model (raster) per la nostra area (1430x1700)
The The digital Forest Mapdigital Forest Map looks reasonable…. looks reasonable….
…….yet some details are quite strange..yet some details are quite strange.
Regions in
Colours
Strata in Yellow lines
Regional approach - FIRS regions and Regional approach - FIRS regions and stratastrata
Similar NDVISimilar NDVI
profiles may haveprofiles may have
a differenta different
meaning in meaning in
different regionsdifferent regions
Black lines FIRS regionsRed lines FIRS strata
AOI per partner:•Yellow: SSC•Green: SC-DLO•Blue: CNRM•Purple: ARCS•Orange: IUAV
Regional approach
Source: EMAP unit (SAI-JRC)
Defined Classification MethodologyMulti-spectral
AVHRR dataMasksNDVI ‘97 CORINE
Pure Pixels
MaskedNDVI
Classification
SignaturesDistances
1st - best2nd - bestclasses
CompilationEnhancements
Post-classification
Regions and strata
Data
base
PELCOM 1km land cover database Conif. Forest
Decid. ForestMixed Forest
Arable landIrrigated landPermanent cropsShrublandBarren landIce and SnowWetlandsWater
Grassland
Urban
PELCOM land coverPELCOM land cover - Pixel: 1.1 kmPixel: 1.1 km
Our AreasAreas of Interest of Interest, each including one or more FIRS strata. The parts pointed at by arrows show where strata extend into regions for which no CORINE info exists.
IUAV - South-East Europe - Final classified imageIUAV - South-East Europe - Final classified image
Scelta delle immagini e problemi connessiScelta delle immagini e problemi connessi
Principal Components Analysis (PCA)Principal Components Analysis (PCA)
• Channel values used as variables are often higly
correlated.A PCAPCA is used to convert the initial variables to new latent variables, called Principal Components, that are uncorrelated. This operation:
• reduces the dimensionality of the description
• filters out noise (i.e., casual, unimportant, non- structural aspects).A PCA is often applied before clustering, to savecomputation time.
AssumptionAssumption
•two pixels are two pixels are similar when they lie when they lie closeclose to each to each otherother
in the feature space (this depends on the way thein the feature space (this depends on the way thedistancedistance is defined). is defined).
To cluster
Create groups (classes) ascompact as possible in thefeature space
Now, a short description Now, a short description
of the main approachesof the main approaches
for clustering...for clustering...
Supervised methodsSupervised methods
• A suitable training set must be defined.
• The target-themes (water, built-up, vegetation, etc.)The target-themes (water, built-up, vegetation, etc.) are given a priori;are given a priori;
• The theme to which any pixel included in the training set belongs must be explicitly stated.
The classifier learns from the training set to recognize the themes’ features, and to assign all new pixels to the most appropriate theme.
ProblemsProblems
• Pixels are generally mix of different themes
• The construction of the training settraining set is a hard and error-prone job.
• gli esempi presentati alla rete consistono solo di un insieme di pattern (vettori di attributi), classificati per similarità.
• Ad ogni nodo (i, j) della mappa è associato un vettore di riferimento m i,j (codebookcodebook)
Kohonen’s MapKohonen’s Map The pixel with pattern xx is assigned to class (1,1)
Reti SOM - procedura di istruzione
• Ciascun pattern xx viene confrontato con tutti i vettori mmijij ed assegnato al nodo associato al
vettore mmijij più somigliante.
• Il nodo vincitore e quelli vicini vengono modificati in modo da accentuare la loro somiglianza con il pattern xx.
• Una SOM sviluppa in modo autonomo durante la fase di apprendimento un’organizzazione interna: pattern di ingresso simili attivano lo stesso nodo della mappa.
Le applicazioni sono di carattere esplorativo
Primi esperimenti di classificazione...Primi esperimenti di classificazione...
• Classificazione non supervisionataClassificazione non supervisionata
Details from the classified image issued by a classifier
trained on selected ares. Built-up areas are split
into two density zones.
Calcolo sistematicoCalcolo sistematico
di distribuzionidi distribuzioni
statistiche spazialistatistiche spaziali
• StatisticheStatistiche sulla distribuzione dei temi calcolate da sulla distribuzione dei temi calcolate da CORINE e dal DTM (quote).CORINE e dal DTM (quote).
• SistematicamenteSistematicamente, per ogni tema ed ogni area., per ogni tema ed ogni area.
• ObiettivoObiettivo: aiutare l’interpretazione delle classi, formulare : aiutare l’interpretazione delle classi, formulare
opportune regole di post-classificaziome opportune regole di post-classificaziome
CORINE ha 43 temi PELCOM ha 11 temi, da suddividere in sub- temi quando possibile
North Italy (ITN) - Distribuzione di temi sulla quotaNorth Italy (ITN) - Distribuzione di temi sulla quota
North Italy - Distribuzione di temi sulla quotaNorth Italy - Distribuzione di temi sulla quota
Italia Centrale - Distribuzione di temi sulla quotaItalia Centrale - Distribuzione di temi sulla quota
Determinazione di regole di decisione (post-classificazione)Determinazione di regole di decisione (post-classificazione)
Altre statisticheAltre statistiche per area e temaper area e tema calcolate durante leclassificazioni: mappe raster della distanza dei pixeldai temi (rappresentati dalla firma spettrale calcolataa partire dai pixel di addestramento)
Strongly methodological, the work concerned mainly:
• the optimal territorial stratification;
• the choice of the images to be used;
• extended experimental comparison of the available clusering methods
Eventually, a supervised supervised method was chosen that included:
• an original way of choosing the training pixels;
• the direct utilization of available ancillary information within the clustering procedure itself.
The ClassificationThe Classification
•Variables: 20 images NDVI during year 1997, produced by DLR, Berlin;
•Training pixels: for each theme, a set of pixels were chosen in each region, that were sufficiently pure in CORINE.
•Ancillary information: CORINE, Digital Terrain Model
at 1 km., Digital Chart of Forests, Bartholomew Maps,
FIRS strata, administrative statistics, etc.
1. Choice of training pixel1. Choice of training pixel
(the method can be easily extended to Neural(the method can be easily extended to Neural
Network Classifiers)Network Classifiers)
Distribution of training pixelsDistribution of training pixels
used for Italy in PELCOM.used for Italy in PELCOM.
For each area of interest,
a certain number of “pure”
pixels (at least 87% belonging
to the same theme in CORINE)
have been selected,
and used to train the classifier.
All other pixels are then
assigned to the most similar
group.
ITN - Projection of candidate training pixelscandidate training pixels onto the first factorplane. For most themes the level of confusion is high.
Confusion matrix for initial training pixelsColumns: pixels theme in CORINE.Rows: theme to which pixels are assigned by the classifier, after computing the themes’ signatures.
Selection of the most reliable training pixelsSelection of the most reliable training pixels(repeated for each region)
Initial selection of all candidates (14/16 pure CORINE)
Computation of the signatures for all relevant themes
Assignment of each training pixel to the closest signature
according to the Mahalanobis distance, thus keeping into
account the dispersion of each theme in the reference space
Exclusion of ill-assigned pixels
Too many ill-classified pixels?
yes
noEND
ITN - Training pixels retained after several iterations. Confusion is much reduced.
Confusion matrix for the finally selected training pixelsColumns: pixels theme in CORINE.Rows: theme to which pixels are assigned by the classifier, aftercomputing the themes signatures.
2. Clustering Method that makes use of 2. Clustering Method that makes use of the available ancillary informationthe available ancillary information
The ancillary info usedThe ancillary info used while clustering:while clustering:
• For each AoI, For each AoI, the the distribution of each distribution of each theme over altitudetheme over altitude
Problems with clusteringProblems with clustering
•At the 1.1 km scale pixels are mix of themes, their spectral signaturesspectral signatures are averages;
•The profiles of some themes are very similar;
A classification purely based upon radiometric data can sometimes lead to unreliable or absurd results:
Example: arable land pixels detected at high elevation, owing to their similarity with bare soil.
How should pixel be assigned?How should pixel be assigned?
• In two steps:1. Purely according to the similarity of their
profiles (with much confusion)… 2. … then applying some suitable
decision rules to correct the results.• Or Knowledge-based Knowledge-based
ClassificationClassification
…making use of some relevant ancillary information during the very clustering process.
The Knowledge-based Classification The Knowledge-based Classification AlgorithmAlgorithm1. Selection of a set of reliable training pixels
2. For each activeactive theme:•computation of the average signature•computation of the specific variance-covariance
matrix and of its inverse (needed to compute Mahalanobis distances)
3. The elevation is divided in user-defined intervals and the table TargetTarget that cross-tabulates themes vs. elevation ranges is computed from CORINE.
4. TargetTarget is corrected by taking into account also contributions to active themes coming from pixels that belong in CORINE to non-active themes.
Current:Current:themes versus elevation as computed from clustering
5.Pixels are assigned to active themes according to the Mahalanobis distance. This is based purely on pixels’ signature.For each pixel, the Mahalanobis distances from all active themes are saved to a temporary file.The table CurrentCurrent of themes’ frequency vs. elevation is computed.
6. At this point we have two matrices:
Target:Target:themes versus elevation as in CORINE
• The classification is changed iteratively so as to have the computed table CurrentCurrent better match the knowledge base (CORINE) represented by TargetTarget;
• at the same time, also TargetTarget is slightly modified so as to keep somehow into account info in input data.
Common structure of matrices TargetTarget and CurrentCurrent:•n user-defined elevation ranges;
•p active themes;
•Cij is the frequency of themej at elevation range i in the CurrentCurrent table;
•on each iteration, assignment to cell (i,j)
- is encouraged if Cij < Tij
- is dis-couraged if Cij > Tij
7. A table of correction factors CFij (also sized n x p) is defined. Initially they are all equal to 1.
Each correction factor refers to one cell. CFij is used to correct the Mahalanobis distance from theme j of pixels at elevation range i.
8. A cicle of iterationscicle of iterations is started, aimed at orienting the assignments (no hard constraints), so as to encourage CurrentCurrent to adhere to TargetTarget.At the same time also TargetTarget is corrected (though less), moving it towards CurrentCurrent.
One iterationOne iteration (cycle 9-10 is repeated)
• re-read its Mahalanobis distances from file, and
correct each with the correction factor CFij
appropriate to the theme and to the pixel’s elevation;
• assign the pixel to the theme for which the so-corrected Mahalanobis distance is minimum.
9. For each pixel::
10. Update the CurrentCurrent distribution Update the correction factors:
Update TargetTarget
All AOIs - 12/16 pure pixels (not used for training)Columns: theme of pixel in CORINE.Rows: theme to which pixel is assigned by the classifier.
V A L I D A T I O N ( 1 )
All AOIs - 12/16 pure pixels (not used for training)Columns: theme of pixel in CORINE.Rows: theme to which pixel is assigned by the classifier.
V A L I D A T I O N ( 2 )
Forests are aggregated, as well as grassland/pastures.
RemarksRemarks
• frequent themes, like arable or deciduous, are well classified;
• less frequent themes, like permanent crops, coniferous or grassland, are better classified in AOIs for which enough training pixels can be found;
• when the classification was attempted with only few training pixels (pastures or permanent crops in some areas), results were unreliable;
An excerpt of Validation tables- summary per themeNon-irrigated arable land
Irrigated arable land + Rice
ITS - Spatial distribution of irrigated land
A3 - Spatial distribution of irrigated land
ConclusionsConclusions• When the target themes are pre-defined, a supervised classifier is convenient.
• Statistical and neural methods lead to results of comparable accuracy.
• The methods used with statistic classifiers in order to select optimally the training pixels, as well as the knowledge-based classification, should be extended to neural classifiers.
The Natural Capital Index (NCI)
enables to monitor and describe
the current and future
State of the Environment
100%
100%0%
quality
quantity
Principle Natural Capital Index:
Quantity: % area
Quality : % baseline state
NCI = quantity x qualityNCI = quantity x quality
•self-regenerating current / natural baseline • man-made current / cultural baseline
} “gap-analysis”
Self-regenerating areas are defined as:
virgin land, nature reserves; all forests except virgin land, nature reserves; all forests except wood plantations with exotic species; areas wood plantations with exotic species; areas with shifting cultivation; all fresh water areas; with shifting cultivation; all fresh water areas; and extensive grasslands (marginal land used and extensive grasslands (marginal land used for grazing by nomadic livestock).for grazing by nomadic livestock).
“ “all not human-dominated land, all not human-dominated land, irrespective of wether it is pristine or irrespective of wether it is pristine or
degraded”degraded”
Man-made Areas are defined as:
arable land; permanent cropland; wood arable land; permanent cropland; wood plantations with exotic species; pasture for plantations with exotic species; pasture for permanent livestock; urban areas; permanent livestock; urban areas; infrastructure; and industrial areas. Most infrastructure; and industrial areas. Most domesticated land is in fact agricultural domesticated land is in fact agricultural land.land.