Multi-temporal satellite imagery and data fusion for improved land cover information extraction

Land use information is one of the most sought inputs for various resource and environmental management studies as well as climate models. In this study, an attempt was made to obtain land cover information from temporal data set of Advance Wide Field Sensor aboard Indian Remote Sensing satellite IRS-P6 using data mining classification technique. This study mainly focuses on the utility of visually interpreted thematic maps as an additional input for improving classification accuracies. The temporal data sets were co-registered to sub-pixel accuracy and were atmospherically corrected using modified dark pixel subtraction method. The visual thematic maps (wastelands and forest cover maps) were also co-registered to satellite data to a near pixel accuracy. Digital values were extracted for various classes and rule sets were generated using See-5 data mining software. These rule sets were ported into ERDAS Imagine Knowledge Engineer and the temporal data sets were classified. The results indicate that temporal satellite data at monthly intervals have a good potential to capture the land cover information including temporal dynamics of crop cover in agricultural lands. Addition of legacy maps obtained through monoscopic visual interpretation has helped to improve classification accuracies significantly. However, there exists a co-registration issue between visual maps as well as satellite data that have influenced the classification accuracies. The decision tree classification algorithm (See-5) used in this study is able to exploit the temporal variation in target spectral properties as well as thematic information from legacy maps satisfactorily. There has been a substantial improvement in various categories of forests as well as wastelands due to addition of visual maps. This has further reduced the misclassification of other vegetation categories, thereby improving the overall classification accuracy. Overall, kappa statistic of 0.885 was achieved with multitemporal satellite data, which was further improved to 0.932 with the addition of visual maps.