Image classification aims to classify pixels into themes or land covers (Lillesand et al, 2014), and is usually achieved by grouping similar pixels together into classes. There are a few ways in which this can be done, two of which are unsupervised and supervised classification.
Unsupervised Classification:
The process of unsupervised classification can be seen in the flow diagram above. It can either be done through a process using K-means or isodata algorithms. In practical 5, I completed an unsupervised classification using the isodata algorithm to classify 6 different land cover types on an image of the Hong Kong harbour. Learning about the individual spectral signatures (see blog 3) provided a greater understanding of how to assess which spectral responses should be combined, as well as what land covers these spectral responses were representing. This will have resulted in a more accurate classification and match to the true colour image.
Supervised Classification:
Supervised classification uses training sites that are identified by the user to classify land cover types. The software uses the spectral signatures of the training sites to identify the land cover type (GISGeography, 2016). In practical 6, we used a supervised classification technique to classify the 6 land cover types on the Hong Kong harbour. Overall, this method produced the better classification, both in visual comparison to the true colour image, and statistically in the accuracy assessment.
Practical examples of using classification:
I found that doing the different classification processes on the same image helped to develop my understanding of the process and visualise the difference between the two methods. From doing these practicals, I questioned how this classification technique was used in research to improve accuracy of remote sensing.
One study used this technique in measuring the effects of wildfires (Joyce et al, 2009). Normally, soil is seen as lighter than vegetation however burn scars contain organic matter and so can appear darker depending on the vegetation (Joyce et al, 2009). This can result in variance and reduce accuracy. To improve this, isodata can be used to classify, with band ratios being able to distinguish burnt and unburnt areas better than individual spectral bands (Joyce et al, 2009).
Supervised classification was also used to map British lowland (Fuller and Parsell, 1990). Training sites were used to map varying land covers in Cambridgeshire, such as: crop types, bare ground (e.g. urban areas), water sites and semi-natural vegetation (Fuller and Parsell, 1990). The overall accuracy of the classification was good, which may have been a result of a using number of training sites. However, the producer’s accuracy and user’s accuracy varied between the land cover types, suggesting that this classification was not uniformly successful. For example, 95% of the land classified as semi-natural, was classified correctly (Fuller and Parsell, 1990). This indicates successful user’s accuracy for this land type. However, less area was classified as deciduous forest than there was in the field (Fuller and Parsell, 1990), indicating a lower producer’s accuracy for deciduous forest land cover in this study.
The results from that study emphasised the importance of completing an accuracy assessment after each classification, to ensure the image can be used with confidence.
References:
Fuller, R.M., and Parsell, R.J. (1990) ‘Classification of TM imagery in the study of land use in lowland Britain: practical considerations for operational use’, International Journal of Remote Sensing, Vol: 11(10), pp. 1901-1917.
GISGeography (2016) Image Classification Techniques in Remote Sensing. Available at: http://gisgeography.com/image-classification-techniques-remote-sensing/
Joyce, K.E., Belliss, S.E., Sergey, V.S, McNeill, S.J., and Glassey, P.J. (2009) ‘A review of the status of satellite remote sensing and image processing techniques for mapping natural hazards and disasters’, Progress in Physical Geography, Vol 33(2), pp. 1-25.
Lillesand, T., Kiefer, R.W., Chipman, J.W. (2004) Remote Sensing and Image Interpretation. (5th Edition). New York: Wiley and Sons.
9321836 