| Abstract | (Summary)
Data were acquired with the Canada Centre for Remote Sensing Convair-580 SAR sensor over selected sites in Thailand in November 1993 as part of the GlobeSAR program. These data were
used to generate RADARSAT simulations of fine-resolution and standard-mode products. Land use mapping and monitoring using SAR data were investigated in the Province of Songkhla in Southern Thailand. A comparison of airborne high-resolution imagery
and simulated RADARSAT data indicated that the reduction in resolution of the simulated products has little impact on the ability to recognize general land use. There is, however, a diminished ability to recognize early rice growth stage because of
the polarization (HH), which seems less sensitive than VV to small surface expression, and because of resolution, as rice paddies are small by nature. Aquaculture features, such as shrimp farms, are quite apparent in the simulated data, meeting the
monitoring requirements for this expanding industry. RADARSAT data should be better suited to compare various rubber plantations as the incidence angle bias will be greatly reduced relative to airborne imagery. Three classification methods were used
to characterize the data. Results showed an important contribution of texture measurements in achieving better classification accuracy. A maximum likelihood classification using land use classes of water, mangrove, open mangrove, young rubber, rubber
plantations, mixed orchards, rice paddies (emergent), and flooded areas shows a 72% accuracy on the entire test site when derived texture measures are used. Forest classes, which include rubber, orchards, and mangrove, are the most difficult to
separate. Their spectral responses are quite wide, producing a somewhat overlapping classification signature. Furthermore, the small plots of land used for rubber plantations and for orchards and their distinct boundaries create problems for the
machine-based classification. However, best separability is achieved using only Chh data and its derived texture measurements. The addition of the CVV data to the classification process reduces the discrimination of these cover types. The difference
in the respective branch structure is better observed with HH-polarized data. Most confusion is easily resolved by considering the geographic context of the plots of land. |