Over the past decade, climate change has become a serious issue, and it has received tremendous attention from investigators in different fields and also from the general public. In order to learn and provide immediate and appropriate actions with respect to climate variations, researchers have proposed various techniques for learning the underlying dynamic behavior in natural environments. In particular, in associated engineering research, convolutional long short-term memory (ConvLSTM) has been utilized to process spatiotemporal observations from the environment and predict the possible future measurements. Although the convolutional layer in this network can capture the local spatial dependencies, the network fails to extract global interactions among hidden features in the date. Moreover, in order to interpret the associated environmental field, a Kriging approach may be utilized in the analysis of the data collected from the monitored area and the estimation of an unknown region, through a Gaussian Process (GP) model. However, the performance of the field mapping is limited by the computationally expensive operation of determining the optimal sensor placement.

To improve the accuracy and efficiency for the indicated goal of environmental monitoring and assessment, the thesis exploits a deep neural network for forecasting an environmental image series and reconstructing an environmental field, based on the acquired environmental data. First, an attention-based ConvLSTM model is developed to perform multi-step image series forecasting. Specifically, a convolutional self-attention (CSA) is designed to learn long-range dependencies within the latent variables during the training procedure. Second, an attention-based deep residual neural network is proposed to speed up the process of selecting the optimal monitoring locations.

The proposed methodologies are evaluated with a real-world dataset. From this evaluation and the obtained experiment results, it is found that the proposed forecasting approach outperforms the existing methods in a reliable and accurate manner. The MSE (mean squared error) of the proposed model is approximately 7% lower than that of the existing methods. Moreover, the proposed model accelerates the process of finding the optimal sensor placements. Specifically, it is shown to achieve approximately 73% improvement in speed and an accuracy of about 90%.