A thesis submitted in partial ful{uFB01}llment of the requirements for the degree of Master of Science

Thesis (M.Sc.) - Asian Institute of Technology, 2000

Remote sensing satellite data have been used to generate the MO}: yield models in various part of the world. As the crop yields are affected by several other factors, such as climatic and soil related variables, generation of such models need to be carried out at local level. A study was carried out to examine the relation between satellite derived Spectral information and soil-related properties with the yield of dry season rice in the central plain of Thailand. Landsat TM data acquired in April 1997 and March 1998, and soil information available at the series level were used to regress with the yield statistics available at sub-district level of the study area. Several yield models containing both Spectral and soil-related variables as predictor variables were generated. The models were applied to predict the yield of dry season rice for the year 1999 by using the spectral information derived from Landsat TM image acquired in February 1999. Digital image classi{uFB01}cation of three sets of images gave quite satisfactory overall classi{uFB01}cation accuracy of 93, 95 and 95 percent for the year 1997, 1998 and 1999, respectively. The class accuracy for dry season rice was 95, 97 and 99 percent, respectively for these images. Since the area is under irrigation, cultivation of second crop of rice is the major land use activity during the dry season, thus easily discriminable with high class accuracies. The areas classi{uFB01}ed as standing rice crop were 40.1, 56.4 and 40.2 percent of the total study area for 1997, 1998 and 1999 images, respectively. Of the 12 independent variables including spectral and soil-related, near-infrared and red bands either alone or in ratio showed higher correlation with rice yield among rest of the Spectral bands in general. Of the soil related variables, organic matter content of the soil showed higher correlation with the yield where as soil depth and soil texture were not found significantly correlated with rice yield. However, in case of 1997 data, a weak correlation was observed between soil texture and yield. The various yield models containing different number of predictor variables were generated. In single year basis, two and three models were generated for the year 1997 and 1998, respectively. For 1997, the final {uFB01}tted model contained two predictor variables and for 1998 three variables. To accommodate the inter-annual variability, both predictor variables and yield of two years were integrated and regression analysis was carried out. The final {uFB01}tted model contained three predicted variables, namely NDVl, OM and ratio 4/3, explaining 61 percent of yield variation. Out of 8 models generated, there were no significant differences between the models so far the prediction power is concerned. The yield of 1999 was forecasted using Model 6, 7 and 8. However, considering that the yield model should be simple, Model 7, which contained NDVI and OM as predictor variables, was relatively superior. This study concluded that that there is correlation between spectra] and soil related variables with yield of dry season rice and which can be explained by their linear relationship to make future forecast. However, some topical recommendations are provided which would help re{uFB01}ning and improving the yield models.