СЕЗОННА ДНАМІКА СПЕКТРАЛЬНХ ХАРАКТЕРСТК ЗЕМНОГО ПОКРВУ ТА ЇЇ РОЛЬ У ДЕШФРУВАННІ ЛІСОВХ НАСАДЖЕНЬ ЗА ЗНІМКАМ LANDSAT



В. В. Миронюк

Анотація


SEASONAL DYNAMICS OF SPECTRAL REFLECTANCE OF LAND
COVERS AND ITS ROLE IN MAPPING FOREST STANDS USING LANDSAT
IMAGES

V. Myronyuk

Introduction. Classification of forests using satellite images is important task for understanding the current state and dynamics of forests. Nowadays there is a number satellite-based platforms that provide remote sensing of Earth with desired spatial and temporal resolution. The Landsat (TM, ETM+, OLI) data is widely used for monitoring of forest resources due to spatial resolution (30 m), wide swath (185 x 185 km) and good radiometric calibration. Time series of satellite images provide repeated observations of the same area, so the best-available-pixels (BAP) could be selected. Mosaicking time series of images tends to improve quality of remote sensing data and reduce an influence of the atmosphere and cloudiness. BAP composited mosaics is proved to be optimal solution for monitoring of large areas, thus methods for composing multidate imagery are needed. Development of precise methods of forest mapping using time series of satellite images is a big challenge for today science.

Analysis of recent research and publication. Due to seasonal changes of vegetation, multidate satellite images are useful for identification of different tree species. Studies conducted on single scenes of Landsat showed that images combined in a form of time series provide the highest precision of tree species classification. The applicability of time series of multispectral images was proved in resent studies by Hansen et al. (2014) for large scale forest mapping. Hill et all. (2010) and Sheeren et all. (2016) used time series for tree species identification, Zald et. all. (2016) investigated the role of seasonal composited mosaics for aboveground biomass modeling. According to Franklin et al. (2015), land cover classification by time series of satellite images is more than 6 % precise compared to single scene classification.

Objective. The objective of the paper is to investigate dynamics of spectral reflectance of different types of land cover using time series of Landsat 8 OLI satellite images.

Methods. We used stratified sampling to collect 4690 randomly distributed points within flat part of Ukraine. The sampling frame was designed on Global Forest Change (GFC) data using FAO recommendations for assessment of maps accuracies. The stratification of the territory based on for GFC data included four classes: 1) stable forest, 2) stable non-forest, 3) forest loss and 4) forest gain.

Google Earth and Bing Maps images of high resolution were used during visual photointerpretation of sampling points. We used three level interpretation scheme so that seven land cover classes were assigned on the first phase. Inside each category several subcategories were selected on second phase. Finally, for forested area the canopy cover was estimated using grid of points overlaid on the Google or Bing image. The number of sampling points was between 205-239 for each region (oblast). As many as 1350 sampling (29%) represented forested area, among them 557 were classified as coniferous forest, 417 – deciduous forest and 276 – mixed forest.

Results. The composited mosaics of Landsat images were created in Google Earth Engine using standard algorithms for composing and cloud removal. We used TOA reflectance images from «LANDSAT/LC8_L1T_TOA» collection. The collection was filtered between 2014-01-01 and 2016-12-31, only images with cloud cover less than 10 % were selected. Thus 621 scenes of Landsat 8 OLI images were selected for analysis that cover flat part of Ukraine.

We used the information from six bands of Landsat images: Band 4 – Red (0.64–0.67 μm) of visible range; Band 5 – NIR (0.85–0.88 μm) of near-infrared range; Band 6 – SWIR 1 and Band 7 – SWIR 2 of short-wave infrared range (1.57–1.65 μm і 2.11–2.29 μm); two thermal bands Band 10 – TIRS 1 and Band 11 – TIRS 2 (10.60–11.19 μm and 11.50–12.51 μm). Blue and green spectra were not used in the analysis since their sensitivity to atmospheric effects. We also calculated following ratio bands: NDVI, Band 4 / Band 6, Band 4 / Band 7, Band 5 / Band 6, Band 5 / Band 7, Band 6 / Band 7. The selected collection of images was grouped by months on per-pixel basis using maximum NDVI value as criteria for selecting pixels. Afterwards we calculated median values for each monthly composited mosaics.

Forest mask increases the accuracy of tree species classification, thus firstly we analyzed the separability of different land covers. Band 6 and Band 7 and their combinations proved to be useful in land cover classification. The band combination Band 5 / Band 7 also has demonstrated high separability of different land cover classes. During a year, there is a trend which is very related to phenology of vegetation. The differences between coniferous and deciduous tree species are obvious in red and near-infrared spectra (Band 4 and Band 5). From October to April when trees are leaf-off, deciduous tree species tend to have higher reflectance in red band and significantly lower – in near-infrared. During May-September this specifics turns over to opposite. The major differences between tree species groups have been captured in short wave-infrared spectra of Band 6 and Band 7. Using band combinations also facilitates tree species classification. It was found in research that thermal bands are not useful for tree species detections.

Discussion. Satellite images acquired during one year or during more durable period could be used for different tasks of land cover classification. The yearly time series informs on phenological changes of vegetation or seasonal stage of land covers like snow, bare soils, vegetation, fallow lands. The composited mosaics from images acquired during several years for some season are important to catch the major changes in land covers. They both could be important for identification of current stage and dynamics of forests. The results of the paper demonstrate the specifics of spectral reflectance dynamics thus important for selecting an appropriate predictor variables in classification models of seasonal Landsat 8 OLI composited mosaics.


Повний текст:

PDF

Посилання


Gomez C. Optical remotely sensed time series data for land cover

classification: A review / C. Gomez, J. C. White, M. A. Wulder // ISPRS

Journal of Photogrammetry and Remote Sensing. – 2016. – Vol. 116. –

P. 55–72.

Good practices for estimating area and assessing accuracy of land change / P. Olofsson, G. M. Foody, M. Herold et all. // Remote Sensing of

Environment. – 2014. – Vol. 148. – P. 42–57.

Hill R. A. Mapping tree species in temperate deciduous woodland using

time-series multi-spectral data / R. A. Hill, A. K. Wilson, M. George,

S. A. Hinsley // Applied Vegetation Science. – 2010. – Vol. 13. (1). – P. 86–99.

Integrating Landsat pixel composites and change metrics with lidar plots to predictively map forest structure and aboveground biomass in

Saskatchewan, Canada / H. S. J. Zald, M. A. Wulder, J. C. White // Remote Sensing of Environment. – 2016. – Vol. 176. – P. 188–201.

Large area mapping of annual land cover dynamics using multi-temporal change detection and classification of Landsat time-series data /

S. E. Franklin, O. S. Ahmed, M. A. Wulder et al. / Canadian Journal of Remote Sensing. – 2015. – Vol. 41. – P. 293–314.

Monitoring conterminous Unite States (CONUS) land cover change with

Web-Enabled Landsat Data (WELD) / M. C. Hansen, A. Egorov,

P. V. Potapov // Remote Sensing of Environment – 2014. – Vol. 140. –

P. 466–484.

Tigges J. Urban vegetation classification: Benefits of multitemporal

RapidEye satellite data / J. Tigges, T. Lakes, P. Hostert // Remote Sensing

Environment. – 2013. – Vol. 136. – P. 66–75.

Tree Species Classification in Temperate Forests Using Formosat-2

Satellite Image Time Series / D. Sheeren, M. Fauvel, V. Josipovic et al. //

Remote Sensing. – 2016. – Vol. 8. – 29 p.


Метрики статей

Завантаження метрик ...

Metrics powered by PLOS ALM

Посилання

  • Поки немає зовнішніх посилань.