Use of local statistics in remote sensing of grasslands and forests
Date
2018-07-04
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Käesolev doktoritöö analüüsib lokaalstatistikute kasutamist rohumaade ja metsade kaugseires. Töö esimene osa käsitleb rohumaade monitoorimist tehisava-radari (synthetic aperture radar (SAR)) abil ning teine osa metsade kaugseiret kasutades optilisi sensoreid. Analüüsides rohumaade niitmise ja C- laineala tehisava-radari interferomeetrilise koherentsuse seoseid leiti, et selle parameetri kasutamisel on potentsiaali niitmise tuvastamise algoritmide ja rakenduste väljaarendamiseks. Tulemused näitavad, et pärast niitmist on VH ja VV polarisatsiooni 12-päeva interferomeetrilise koherentsuse mediaan väärtused statistiliselt oluliselt kõrgemad võrreldes niitmise eelse olukorraga. Koherentsus on seda kõrgem, mida väiksem on ajaline vahe niitmise ja pärast seda üles võetud esimese interferomeetrilise mõõtmise vahel. Hommikune kaste, sademed, põllutööde teostamine, näiteks külv või kündmine, kõrgelt niitmine ja kiire rohu kasv pärast niitmist vähendavad koherentsust ja raskendavad niitmise sündmuste eristamist. Selleks, et eelpoolnimetatud mõjusid leevendada tuleks tulevikus uurida 6-päeva koherentsuse ja niitmise sündmuste vahelisi seoseid. Käesolevas doktoritöös esitatud tulemused loovad siiski tugeva aluse edasisteks uuringuteks ja arendusteks eesmärgiga võtta C-laineala tehisava-radari andmed niitmise tuvastamisel ka praktikas kasutusele. Lisaks näidati, et ortofotodel põhinevate metsa kaugseire hinnangute andmisel on abi lokaalstatistikute kasutamisest. Analüüsides kaugseire hinnangut riigimetsa takseerandmete (national forest inventory) kohta leiti, et näidistel põhinev järeldamine (case-based reasoning (CBR)) sobib hästi selliste kaugseire ülesannete empiirilisteks lahendusteks, kus sisendandmetena on kasutatavad väga paljud erinevad andmeallikad. Leiti, et klasteranalüüsi saab kasutada kaugseire tunnuste eelvaliku meetodina. Võrreldes erinevaid tekstuuri statistikuid näidati, et lokaalselt arvutatud keskväärtus on kõige väärtuslikum tunnus. Järeldati, et nii statistiliste kui ka struktuursete lokaalstatistikute kasutamisega saab lisada pikslipõhistele kaugseire hinnangutele olulist andmestikku.
This thesis studies approaches for remote sensing of grasslands and forests based on local statistics. The first part of the thesis focuses on monitoring of grasslands with SAR and the second part to monitoring of forests with optical sensors. It is shown that there is potential to develop mowing detection algorithms and applications using C-band SAR temporal interferometric coherence. The results demonstrate that after a mowing event, median VH and VV polarisation 12-day interferometric coherence values are statistically significantly higher than those from before the event. The sooner after the mowing event the first interferometric acquisition is taken, the higher the coherence. Morning dew, precipitation, farming activities, such as sowing or ploughing, high residual straws after the cut and rapid growth of grass are causing the coherence to decrease and impede the distinction of a mowing event. In the future, six-day interferometric coherence should also be analysed in relation to mowing events to alleviate some of these factors. Nevertheless, the results presented in this thesis offer a strong basis for further research and development activities towards the practical use of spaceborne C-band SAR data for mowing detection. Further, it was shown that local statistics can be useful for estimation of forest parameters from ortophotos and they could also provide helpful ancillary information to conduct a photo-interpretation tasks over forested areas. It was demonstrated that cluster analysis can be used as pre-selection method for the reduction of remote sensing features. Additionally, it was shown that case-based reasoning (a machine learning method) is well suited for empirical solutions of remote sensing tasks where there are many different data sources available. It was concluded that the use of local statistics adds valuable data to pixel-based remote sensing estimations.
This thesis studies approaches for remote sensing of grasslands and forests based on local statistics. The first part of the thesis focuses on monitoring of grasslands with SAR and the second part to monitoring of forests with optical sensors. It is shown that there is potential to develop mowing detection algorithms and applications using C-band SAR temporal interferometric coherence. The results demonstrate that after a mowing event, median VH and VV polarisation 12-day interferometric coherence values are statistically significantly higher than those from before the event. The sooner after the mowing event the first interferometric acquisition is taken, the higher the coherence. Morning dew, precipitation, farming activities, such as sowing or ploughing, high residual straws after the cut and rapid growth of grass are causing the coherence to decrease and impede the distinction of a mowing event. In the future, six-day interferometric coherence should also be analysed in relation to mowing events to alleviate some of these factors. Nevertheless, the results presented in this thesis offer a strong basis for further research and development activities towards the practical use of spaceborne C-band SAR data for mowing detection. Further, it was shown that local statistics can be useful for estimation of forest parameters from ortophotos and they could also provide helpful ancillary information to conduct a photo-interpretation tasks over forested areas. It was demonstrated that cluster analysis can be used as pre-selection method for the reduction of remote sensing features. Additionally, it was shown that case-based reasoning (a machine learning method) is well suited for empirical solutions of remote sensing tasks where there are many different data sources available. It was concluded that the use of local statistics adds valuable data to pixel-based remote sensing estimations.
Description
Väitekirja elektrooniline versioon ei sisalda publikatsioone
Keywords
grasslands, forests, remote sensing, image analysis, statistical analysis