Abstract
This paper presents a monitoring method for paddy fields with hyperspectral remote sensing images in West Java, Indonesia. The statistical modeling method called sparse reguralization is introduced in two forms, that is, LASSO regression for the rice yield estimation and sparse discriminant analysis for the growth stage classification of rice plants, in order to take advantages of the detailed reflectance spectrum measured by numerous bands and to overcome the difficulties in hyperspectral image analysis such as model overfitting. Results of the experiment with airborne hyperspectral images measured by HyMap indicate that sparse regularization can predict paddy conditions with higher degree of accuracy than several estimation methods commonly used in remote sensing applications, such as normalized difference spectral index, partial least squares, or support vector machines. Besides, the prediction models have a limited number of bands which are expected to be informative to figure out the rice growth situation. The overall error between predicted rice yield of the target area and agricultural statistics is 6.40 %, showing the potential effectiveness of methods described in this paper.
