Deep learning approaches on Landsat observations: a review and meta-analysis

Although multiple summary reviews of deep learning in remote sensing have been conducted, the field currently lacks a comprehensive analysis of deep learning utilizing Landsat observations in single time pixel-based classification, both qualitatively and quantitatively. This study addresses this gap by conducting a systematic review based on 352 case studies extracted from 279 peer-reviewed articles that supported direct comparisons across deep learning and traditional classifiers. Qualitatively, our findings are as follows: (i) Two dimensional Convolutional Neural Networks (CNNs) remain the dominant methods; however, exploring other architectures is becoming a growing trend. (ii) Most studies focused on single countries (e.g., China, USA, and India) and the last decade, (iii) Sentinel-2 are the most commonly fused sensors, and vegetation indices are the most commonly used spectral indices. The quantitative meta-analysis revealed that Deep Neural Networks (DNNs) offer an approximate 5% improvement for non-DNN accuracy of 80%, with potentially higher improvements for more challenging tasks. By contrasting between popular DNN methods, UNets achieve higher overall accuracy (OA) than CNNs (approximately 10% UNet improvement when CNN OA is at 80%). However, UNets are more challenging to implement because they require labels for large contiguous reference patches, while CNNs, at least at 30m Landsat pixel, pose lower spatial requirements. In another comparison, Transformers achieve higher OA than CNNs (approximately 5% Transformer improvement when CNN OA is at 80%), although the accuracy improvements are smaller compared to the increases associated with UNet benefits. With greater flexibility and more standardized architecture, Transformers can potentially simplify the design process of models. However, Transformers also face challenges having higher computational resource and memory requirements. Increasing availability of reference datasets, further algorithmic developments in the machine learning field, and improvements in cloud-based computing power and data storage, could increase the utility of deep learning methods on Landsat observations.