Improved Landsat Operational Land Imager (OLI) Cloud and Shadow Detection with the Learning Attention Network Algorithm (LANA)

Landsat cloud and cloud shadow detection has a long heritage based on the application of empirical spectral tests to single image pixels, including the Landsat product Fmask algorithm, which uses spectral tests applied to optical and thermal bands to detect clouds and uses the sun-sensor-cloud geometry to detect shadows. Since the Fmask was developed, convolutional neural network (CNN) algorithms, and in particular U-Net algorithms (a type of CNN with a U-shaped network structure), have been developed and are applied to pixels in square patches to take advantage of both spatial and spectral information. The purpose of this study was to develop and assess a new U-Net algorithm that classifies Landsat 8/9 Operational Land Imager (OLI) pixels with higher accuracy than the Fmask algorithm. The algorithm, termed the Learning Attention Network Algorithm (LANA), is a form of U-Net but with an additional attention mechanism (a type of network structure) that, unlike conventional U-Net, uses more spatial pixel information across each image patch. The LANA was trained using 16,861 512 × 512 30 m pixel annotated Landsat 8 OLI patches extracted from 27 images and 69 image subsets that are publicly available and have been used by others for cloud mask algorithm development and assessment. The annotated data were manually refined to improve the annotation and were supplemented with another four annotated images selected to include clear, completely cloudy, and developed land images. The LANA classifies image pixels as either clear, thin cloud, cloud, or cloud shadow. To evaluate the classification accuracy, five annotated Landsat 8 OLI images (composed of >205 million 30 m pixels) were classified, and the results compared with the Fmask and a publicly available U-Net model (U-Net Wieland). The LANA had a 78% overall classification accuracy considering cloud, thin cloud, cloud shadow, and clear classes. As the LANA, Fmask, and U-Net Wieland algorithms have different class legends, their classification results were harmonized to the same three common classes: cloud, cloud shadow, and clear. Considering these three classes, the LANA had the highest (89%) overall accuracy, followed by Fmask (86%), and then U-Net Wieland (85%). The LANA had the highest F1-scores for cloud (0.92), cloud shadow (0.57), and clear (0.89), and the other two algorithms had lower F1-scores, particularly for cloud (Fmask 0.90, U-Net Wieland 0.88) and cloud shadow (Fmask 0.45, U-Net Wieland 0.52). In addition, a time-series evaluation was undertaken to examine the prevalence of undetected clouds and cloud shadows (i.e., omission errors). The band-specific temporal smoothness index (TSIλ) was applied to a year of Landsat 8 OLI surface reflectance observations after discarding pixel observations labelled as cloud or cloud shadow. This was undertaken independently at each gridded pixel location in four 5000 × 5000 30 m pixel Landsat analysis-ready data (ARD) tiles. The TSIλ results broadly reflected the classification accuracy results and indicated that the LANA had the smallest cloud and cloud shadow omission errors, whereas the Fmask had the greatest cloud omission error and the second greatest cloud shadow omission error. Detailed visual examination, true color image examples and classification results are included and confirm these findings. The TSIλ results also highlight the need for algorithm developers to undertake product quality assessment in addition to accuracy assessment. The LANA model, training and evaluation data, and application codes are publicly available for other researchers.