Portnoi Mickaël. 2016. Methods for sugarcane harvest detection using polarimetric SAR. Stellenbosch : Stellenbosch University, 129 p. Thesis MSc : Geo-informatics : Stellenbosch University
![]()
|
Version publiée
- Anglais
Utilisation soumise à autorisation de l'auteur ou du Cirad. ID589968.pdf Télécharger (4MB) | Prévisualisation |
Encadrement : Kemp, Jaco ; Todoroff, Pierre
Résumé : Remote sensing has long been used as a method for crop harvest monitoring and harvest classification. Harvest monitoring is necessary for the planning of and prompting of effective agricultural practices. Traditionally sugarcane harvest monitoring and classification within the realm of remote sensing is performed with the use of optical data. However, when monitoring sugarcane, the growth period of the crop requires a complete set of multi-temporal image acquisitions throughout the year. Due to the limitations associated with optical sensors, the use of all weather, daylight independent Synthetic Aperture Radar (SAR) sensors is required. The added polarimetric information associated with fully polarimetric SAR sensors result in complex datasets which are expensive to acquire. It is therefore important to assess the benefits of using a fully polarimetric dataset for sugarcane harvest monitoring as opposed to a dual polarimetric dataset. The dual polarimetric dataset which is less complex in nature and can be acquired at a fee much less than that of the fully polarimetric dataset. This thesis undertakes the task of identifying the value of fully polarimetric data for sugarcane harvest identification and classification. Two main experiments were designed in order to complete the task. The experiments make use of fully polarimetric RADARSAT-2 C-band imagery covering the southern part of Reunion Island. Experiment 1 made use of a multi temporal single feature differencing technique for sugarcane harvest identification. Polarimetric decompositions were extracted from the fully polarimetric data and used along with the inherent SAR features. The accuracy with which each SAR feature was able to predict the sugarcane harvest date for each field was assessed. The polarimetric decompositions were superior in classification accuracy to the inherent SAR features. The Van Zyl volume decomposition component achieved an accuracy of 88.33% whereas the inherent SAR backscatter feature (HV) achieved an accuracy of 80%. Hereby displaying the value of the added information associated with fully polarimetric SAR data. The SAR backscatter channels did not achieve accuracies as high as the polarimetric features but did display promise for single feature sugarcane harvest identification when using only a dual polarimetric dataset. Experiment 2 assessed six different machine learning classifiers, applied to single-date, dual- and fully polarized imagery, to determine appropriate combinations of machine learning classifier and SAR features. Polarimetric decompositions were extracted from the fully polarimetric data and mean texture measures were then calculated for all SAR features for both the dual- and full polatrimetric data. A multi-tiered feature reduction method was undertaken in order to reduce dataset dimensionality for the dual- and fully polarised datasets. In general, the reduction in features resulted in improved accuracies. The best sugarcane harvest accuracy was achieved using the Maximum likelihood classifier using on the HV and VV backscatter channels (96.18%). The results from Experiments 1 and 2 indicate that SAR C-band data is suitable for sugarcane harvest monitoring and mapping in a tropical region where optical data have limitations associated with cloud cover and large amounts of moisture in the atmosphere. With the availability of dual polarised Sentinel-1 SAR data, future research should be focussed on the use of a dual polarimetric sugarcane harvest monitoring tool and should be extended to focus not only on sugarcane but other crops which contribute largely to the agriculture and economic sectors.
Résumé (autre langue) : Afstandswaarneming word lankal reeds gebruik as 'n metode in die monitering van die oes van gewasse asook vir oes-klassifikasie. Oes-monitering is nodig vir die beplanning en stimulering van effektiewe landboupraktyke. Tradisioneel word suikerriet oes-monitering en klassifisering, binne die raamwerk van afstandswaarneming, uitgevoer met die gebruik van optiese data. Tog, met die monitering van suikerriet, vereis die groeiperiode van die gewas 'n volledige stel multi-temporale beeldverwerwings dwarsdeur die jaar. As gevolg van die beperkings geassosieer met optiese sensors, word die gebruik van daglig onafhanklike sintetiese gaatjie radar sensors, eerder bekend as Sintetiese Apertuur Radar (SAR) sensors, vir gebruik in alle weersomstandighede, vereis. Die bykomende polarimetriese informasie geassosieer met ten volle gepolarimetriese SAR sensors lei tot komplekse datastelle wat duur is om aan te skaf. Dit is daarom belangrik om die voordele van die gebruik van 'n ten volle gepolarimetriese datastel vir suikerriet oes-monitering in teenstelling met 'n tweeledige polarimetriese datastel wat minder kompleks van aard is en teen 'n fooi veel minder as dié van die ten volle gepolarimetriese datastel verkry kan word, te evalueer. Hierdie tesis onderneem die taak van die identifisering van die waarde van ten volle gepolarimetriese data vir suikerriet oes-identifikasie en -klassifikasie. Twee hoof-eksperimente is ontwerp om die taak te voltooi. Die eksperimente gebruik ten volle gepolarimetriese RADARSAT-2 C-band beelde wat die suidelike deel van Reunion-eiland dek. Met eksperiment 1 is gebruik gemaak van 'n multi-temporale enkelkenmerk differensie- tegniek vir suikerriet oes-identifisering. Polarimetriese ontledings is uit die ten volle gepolarimetriese data geneem en saam met die inherente SAR kenmerke gebruik. Die akkuraatheid waarmee elke SAR kenmerk in staat was om die suikerriet oes-datum vir elke veld te voorspel, is geëvalueer. Die polarimetriese ontledings was beter in klassifikasie- akkuraatheid as die inherente SAR kenmerke. Hiermee word die waarde van die bykomende inligting geassosieer met ten volle gepolarimetriese SAR data, geopenbaar. Die SAR teruguitsaaiingskanale het nie akkuraathede so hoog soos die polarimetriese kenmerke bereik nie, maar het belofte getoon vir enkelkenmerk suikerriet oes-identifikasie wanneer slegs van 'n tweeledige polarimetriese datastel gebruik gemaak word. Met eksperiment 2 is ses verskillende masjien-leer klassifiseerders, toegepas op enkeldatum, tweeledige en ten volle gepolariseerde beelde, geëvalueer om toepaslike kombinasies van masjien-leer klassifiseerder en SAR kenmerke te bepaal. Polarimetriese ontledings is geneem uit die ten volle gepolarimetriese data en beteken dat tekstuur afmetings toe bereken is vir alle SAR kenmerke vir beide die tweeledige- en ten volle gepolarimetriese data. 'n Multi-reeks kenmerkreduksie-metode is onderneem om datasteldimensionaliteit te verminder vir die tweeledige- en ten volle gepolariseerde datastelle. Oor die algemeen het die redusering van kenmerke verbeterde akkuraatheid tot gevolg gehad. Die beste suikerriet oes-akkuraatheid is behaal deur die Maksimum waarskynlikheid klassifiseerder met behulp van die HV en VV teruguitsaaiingskanale (96,18%) te gebruik. Die resultate van eksperimente 1 en 2 dui daarop dat SAR C-band data geskik is vir suikerriet oes- monitering en kartering in 'n tropiese streek waar optiese data beperkings toon wat geassosieer word met wolkbedekking en groot hoeveelhede vog in die atmosfeer. Met die beskikbaarheid van tweeledige gepolariseerde Sentinel-1 SAR data, behoort toekomstige navorsing gefokus te wees op die gebruik van 'n tweeledige polarimetriese suikerriet oes- moniteringshulpmiddel en behoort dit uitgebrei te word om te fokus nie net op suikerriet nie, maar ook ander gewasse wat grootliks bydra tot die landbou- en ekonomiese sektore.
Mots-clés libres : Harvest identification, SAR, RADARSAT-2, Machine learning, Polarimetry
Auteurs et affiliations
- Portnoi Mickaël, Stellenbosch University (ZAF)
Source : Cirad-Agritrop (https://agritrop.cirad.fr/589968/)
[ Page générée et mise en cache le 2023-08-29 ]