The accessibility of high spatial and high temporal satellite imagery time series has enabled the development of methods of multitemporal land cover classification, which should improve the quality of classification due to time information. The dissertation examines two multitemporal classifications: quasi-multitemporal classification and time series based classification. In the first one, image time series are used as the attributes of single-date classification, while the second one compares the development of certain spectral characteristics over time. The classifications of five basic land cover classes (forest, grass, arable land, water, build-up) and six basic crops (maize, wheat, barley, pumpkin, rapeseed, triticale) are performed using various input images, attributes, mapping units and images of different sensors. Emphasis is given to segmentation processes since the segments as basic units of multitemporal classification are not well researched globally. In addition, several possible effects on the classification result are analysed in detail to provide guidelines for obtaining high precision in short processing time. The results show that satellite image acquisition time, besides spectral values, is the most important attribute in the classification. The quasi-multitemporal classification returns a much higher overall accuracy (+8% basic classes, +16% crops), with an average total accuracy of 90% (basic classes) and 88% (crops) giving the useful operational value. The results of the time series based classification are worse (-1% basic classes, -25% crops), the processing time being extremely long, which makes the method as currently unacceptable for practical use. An important finding of this dissertation is that segments are not the most suitable mapping units for the multitemporal classification. Regardless of the production process, segments give worse classification accuracy than reference polygons (-5% basic classes, -18% crops) and pixels (-5% basic classes, -16% crops).
|
