Deep learning for land cover classification and environmental analysis using high-resolution remote sensing data

Deep Learning (DL) has become a breakthrough technology in machine learning, and opportunities are emerging for applications in the Earth Observation Science. Originally, DL algorithms were developed for computer vision problems, and the feasibility of these models needs to be explored for remote sensing topics, such as land cover mapping. Most DL studies are focused on urban mapping or a single scene, and the classification framework needs to be discussed for multiple-image, large-area implementation using high spatial resolution data. In this dissertation, three studies were conducted to explore DL algorithms in different contexts: (i) development of new multi-scale object-based convolutional neural network (multi-OCNN) for large-area land cover classification at 1-m spatial resolution; (ii) evaluation of the deep neural network (DNN) and WorldView-3 image for small wetland mapping; and (iii) mapping of structural conservation practices in Midwest U.S. cropland areas using semantic segmentation method. The research gaps of these studies are further explained in each chapter. In the first study, our findings show that combination of image segmentation, object analysis, and multiscale CNNs in the new multi-OCNN method achieved higher accuracy and faster classification compared to those results from pixel-based CNN and fixed-OCNN. The object analysis allows the selection of convolutional locations and appropriate input window for CNN prediction, which reduces the number of predictions per object and improves the spatial agreement of object and input patch size. Finally, the results show that multi-OCNN approach is a practical alternative for large-area application using traditional CNNs. In the second study, a pixel-based DNN classification produced a fine-detail mapping of wetlands in the Millrace Flats Wildlife Management Area, Iowa, USA. Our results show that DNN model achieved a good classification performance (0.933) in such a complex area, and the results are quite similar to other machine learning methods (random forest, support vector machine and k-nearest neighbor). The results illustrate the benefits of feature selection procedure in the model performance, and the combination of spectral and topographic-related metrics is recommended. Also, this study demonstrated the impact of spatial-resolution on wetland classification, where the agreement between classified and reference areas increased from 61.22% at 30-m to 90.36% at 1.2-m resolution. In the last study, the adapted U-Net architecture was implemented to classify structural conservation practices (SCP) across Midwest U.S. croplands (overall accuracy: 76.8%). In general, the states with the highest percentage of SCPs in cropland areas are Iowa (26%), Illinois (15%), and Nebraska (11%) of total area (6,642 km2). The spatial distribution of SCPs shows the largest occurrence in the southwestern Iowa and eastern Nebraska. Our findings show that occurrence of percentage of SCP in cropland area is partially associated with soil and topographic characteristics such as slope and saturated hydraulic conductivity. In addition, most regions with high soil erosion rates present the largest percentage of SCP areas in croplands as well, indicating conservation efforts by farmers. The development of this product has positive implications for conservation programs, and geospatial inventory is the easily accessible product for large-area evaluation of conservation practices across Midwest U.S. croplands. In conclusion, this dissertation explores the potential of DL algorithms for different classification problems using high spatial resolution imagery, and remote sensing users can benefit from insights of each study