Evaluations of Deep Learning Methods for Land Cover Classification of Satellite Imagery

Deep learning (DL) has rapidly emerged as a highly effective approach in remote sensing image analysis, due to its ability to automatically extract complex spatial and spectral patterns from large volumes of satellite data. Leveraging architectures such as convolutional neural networks (CNNs), DL models have shown advantages over traditional classification and machine learning methods in key applications like land cover mapping, object detection, and time-series classification, making them crucial tools for modern Earth observation research. Recent advancements in DL have further improved the DL model generalization and scalability, demonstrating significant potential for more applications at regional scales. However, limited studies have performed detailed assessment of how DL models can be scaled up and perform at regional scales, especially in the diverse and complex landscapes of sub-Saharan Africa.West Africa has become a hotspot of rapid population growth that has led to significant changes in land use and land cover. Agricultural expansion into natural vegetation has accelerated the changes in the human-modified landscape, specifically in Senegal. Land cover mapping using medium-resolution satellite imagery in this complex landscape has struggled to characterize the mosaic of small agricultural fields. The recent increase in availability of very-high-spatial resolution (VHR) satellite imagery (WorldView-2,-3 and PlanetScope), has opened the opportunities to observe and analyze highly detailed land cover changes, including smallholder agricultural systems. The overall aim of this dissertation is to evaluate (i) DL approaches and strategies to overcome the challenges associated with scaling up land cover and croplands mapping, utilizing extensive volumes of very-high-resolution (VHR) data, (ii) employing medium-resolution satellite image time series (SITS), and (iii) using GAN’s to generate additional VHR data from medium resolution data, in Senegal. More specifically, this study first evaluated various training strategies and DL approaches to scale up (increase training data volume and representativeness) fully supervised training, transfer learning, in order to create regional VHR land cover map at 2 m spatial resolution. It also addressed the importance of training data representativeness and accuracy in scaling up DL approaches in relation to DL model prediction accuracy/generalization across diverse study regions. Second, the study introduced a novel DL architecture, namely Dense Predictive Coding - UNet (DPC-UNet) to map croplands from medium resolution satellite imagery time series (SITS) and compare to other established SITS DL models. The developed data framework allowed the SITS DL models to accurately classify multi-year croplands and produce 30 m regional cropland maps. In the third phase of the study, the complementary strengths of both VHR and SITS data were integrated through the use of the Generative Adversarial Network (GAN) framework. This approach enabled the synthesis of realistic VHR imagery, effectively addressing the temporal gaps in VHR WV data availability, by significantly improving the continuity and regular coverage of VHR observations, thus allowing wall-to-wall mapping efforts at 2 m resolution. Overall the dissertation demonstrated the crucial role DL can and will play in satellite image analysis and earth observation applications.