In the field of scientific computing, Fortran programs contain extensive numerical operations based on the finite precision representation of floating-point numbers, making them prone to potential runtime errors, such as arithmetic overflow, division by zero, etc. The key to check for those runtime errors is to infer value ranges of floating-point variables and expressions. In this paper, we propose to conduct value analysis of floating-point programs by combining interval and zonotopic abstractions. First, we introduce a lightweight zonotopic abstraction for floating-point programs, which soundly handles floating-point arithmetic operations but without introducing too many noise terms. On particular, we present a boundary search based algorithm to improve the division operation of the classic zonotopic abstraction. On this basis, we propose to combine interval and zonotopic abstractions to improve the analysis precision. Finally, we developed a static analyzer based on the above approach for floating-point Fortran programs. Experimental results demonstrate that the proposed method enhances the value analysis efficiency of floating-point Fortran programs while maintaining soundness.