This paper describes Japanese-English translation of weather reports by ISOBAR. Recently, there have been developed a lot of automatic translation systems, either operational or experimental. In the domain of meteorological reports, TAUM-METEO (English-French) is the best-known operational system, which is classified as one of the transfer systems. Our translation by ISOBAR is still experimental, but its most remarkable feature is that target language (TL) sentences are generated only from interlingual meaning representations of input source language (SL) sentences. That is, our system has no transfer component. In this approach, one sentence can be translated into plural TL sentences which are in the paraphrase relation. During semantic processing of SL sentences, ISOBAR, unlike Wilks' model, is designed to reject definitely such sentences as are semantically anomalous or contradictory to its knowledge. We discuss several merits and demerits of this approach, and show some experimental results in ISOBAR.