Classification of multi-temporal SPOT-XS satellite data for mapping rice fields on a West African floodplain
The use of satellite imagery data to accurately map agricultural land use in semi-arid Africa has proven difficult due to the small size of cropped fields, low vegetative cover within and outside cropped areas, and high correlation between the seasonal trajectories in the signatures of cropped and uncropped surfaces. This paper evaluates the potential of computer-aided classification of multi-temporal imagery data to map traditionally-managed rice fields in Sahelian West Africa. SPOT-XS imagery data covering the southwestern corner of the Inland Niger Delta of Mali were acquired on three dates in 1988. These dates encompass the pre-flood period when ploughed surfaces contrast strongly with surrounding land surface and the first half of the flood period. Two alternative classification approaches were evaluated. Supervised and unsupervised classification of a database formed by principal component reduction proved unsuccessful as a result of the high spectral heterogeneity of ploughed and unploughed surfaces. An alternative classification approach was developed which utilizes a series of steps (unsupervised classification, stratification, and supervised classification) that progressively reduce image heterogeneity in such a way to better aid interpretation and comparison with training data. Seventy-one per cent of ploughed-field reference sites were classified correctly using this approach, while 8-20 per cent of reference sites for cover types with a potential for being confused with ploughed fields were misclassified as ploughed fields. A fuller evaluation was performed using a posterior accuracy assessment of classifications produced at five intermediate stages of the procedure. The suitability of the procedure for land-use classification in spectrally-heterogenous areas is discussed.
No Reference information available - sign in for access.
No Citation information available - sign in for access.
No Supplementary Data.