Extracting oilseed rape growing regions based on variation characteristics of red edge position

Abstract: The accurate extraction of oilseed rape growing regions is the premise of growth monitoring, yield estimation and disaster assessment, and remote sensing has been proved to be an effective way for this task. Hyperspectral remote sensing has the features of many more bands, higher spectral resolution, being rich in information and so on, which provides new technical means for extracting oilseed rape growing regions. Changxing, the county with the largest oilseed rape cultivated area in Zhejiang Province was selected to be the study area in this paper, and two EO-1 Hyperion images acquired on April 4th and May 6th, 2004 were adopted, corresponding to the full-bloom stage and pod stage respectively for oilseed rape growing in this region. After detailedly preprocessing of L1R data, the "linear four-point interpolation" method was adopted to get the red edge position (REP) of both stages. REP statistical histograms of typical oilseed rape growing regions and woodland covering both stages were generated and analyzed. The result showed that for both oilseed rape and woodland, the histograms spanning the two stages didn't have overlaps, and for oilseed rape, the REP value demonstrated an obviously "blue shift" characteristic from April 4th to May 6th. On the contrary, the REP value of woodland demonstrated clearly "red shift" characteristic during the same period, which was distinct from oilseed rape. Ignoring other over wintering crops like winter wheat because of the small amount of planting in Changxing that year, and according to the unique "blue shift" characteristic for oilseed rape from full-bloom stage to pod stage, differing from other vegetations, a "decision tree" was built, containing the algorithms of eliminating non-vegetation areas, non-oilseed rape growing regions, and pseudo-growing regions. Then oilseed rape growing regions were extracted based on this method. Next, the result was verified through 500 random sampling points combined with a visual interpretation method, the points were generated within the common coverage of both images, and among them 124 points were located in oilseed rape growing regions. The verification was performed employing an ENVI-ROI tool, and the result showed that the omission error and commission error were 13.7% and 15.7% respectively, total accuracy was 92.6%, and the Kappa coefficient reached up to 0.803, all above indicating that the research had achieved good results. The method provided by this paper takes full advantage of hyperspectral remote sensing, which can retrieve red edge parameters of vegetation with only a few bands, and gets rid of traditional information extracting methods like spectral match. Moreover, the accuracy is not sensitive to spectral differences of surface features, so it has every qualification to be a set of new ideas and solutions for extraction of oilseed rape growing regions by remote sensing.