The objective of this study is to investigate a novel temperature and humidity prediction algorithm for smart greenhouse based on the machine learning method. The smart greenhouse is known to increase farm production by automatically controlling temperature and humidity and other factors. However, maintaining constant inside temperature and humidity in the conventional smart greenhouse system is still a problem because of the multiple time delay elements. To solve the problems, prediction control scheme is required. But, since the system is highly nonlinear with the lack of sensory data, predicting accurate temperature and humidity is very challenging. In this paper, the multi-dimensional Long Short-Term Memory networks (LSTMs) is being applied to deal with the unstructured greenhouse environmental data. The designed LSTMs learning model is trained with the 27 dimensional data which comprises of all the greenhouse control parameter and environmental sensory data. The prediction performance was evaluated using the short, mid and long term experiments. Also, the comparison with the conventional recurrent neural networks (RNNs) based prediction algorithm was done using the experimental results and later on discussions.