Large Area Land Cover Mapping Using Deep Neural Networks and Landsat Time-Series Observations

This dissertation focuses on analysis and implementation of deep learning methodologies in the field of remote sensing to enhance land cover classification accuracy, which has important applications in many areas of environmental planning and natural resources management. The first manuscript conducted a land cover analysis on 26 Landsat scenes in the United States by considering six classifier variants. An extensive grid search was conducted to optimize classifier parameters using only the spectral components of each pixel. Results showed no gain in using deep networks by using only spectral components over conventional classifiers, possibly due to the small reference sample size and richness of features. The effect of changing training data size, class distribution, or scene heterogeneity were also studied and we found all of them having significant effect on classifier accuracy. The second manuscript reviewed 103 research papers on the application of deep learning methodologies in remote sensing, with emphasis on per-pixel classification of mono-temporal data and utilizing spectral and spatial data dimensions. A meta-analysis quantified deep network architecture improvement over selected convolutional classifiers. The effect of network size, learning methodology, input data dimensionality and training data size were also studied, with deep models providing enhanced performance over conventional one using spectral and spatial data. The analysis found that input dataset was a major limitation and available datasets have already been utilized to their maximum capacity. The third manuscript described the steps to build the full environment for dataset generation based on Landsat time-series data using spectral, spatial, and temporal information available for each pixel. A large dataset containing one sample block from each of 84 ecoregions in the conterminous United States (CONUS) was created and then processed by a hybrid convolutional+recurrent deep network, and the network structure was optimized with thousands of simulations. The developed model achieved an overall accuracy of 98% on the test dataset. Also, the model was evaluated for its overall and per-class performance under different conditions, including individual blocks, individual or combined Landsat sensors, and different sequence lengths. The analysis found that although the deep model performance per each block is superior to other candidates, the per block performance still varies considerably from block to block. This suggests extending the work by model fine-tuning for local areas. The analysis also found that including more time stamps or combining different Landsat sensor observations in the model input significantly enhances the model performance