Harvesting the Landsat archive for land cover land use classification using deep neural networks: Comparison with traditional classifiers and multi-sensor benefits

Abstract

The Landsat archive, with a multi-decadal global coverage is a prime candidate for deep learning classification methods due to the large data volume. Local studies have evaluated deep learning methods on Landsat observations. However, these models often saturate at high accuracies due to limited reference dataset size thus do not fully explore the potential of deep classifiers. Furthermore, no provisions are taken to investigate algorithmic performance of challenging classification areas. To address these shortcomings in this research, Landsat 5, 7 and 8 observations were combined within the continental United States to create one of the largest to date reference dataset containing about 21 million labeled annual temporal sequences. Difficult to classify reference samples were isolated by examining labelsin the immediate vicinity. Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) deep learners were integrated to capture temporal and spatial relationships, respectively. Classification mapping accuracy was contrasted with a commonly implemented large-scale mapping method, the Random Forest (RF).Results indicate substantial classification improvements of deep learning methods (DLMs) over the RF. These improvements are more pronounced on challenging to classify pixels in heterogenous areas. RF classification accuracy reaches about 70% on average, while DLMs are at 86%-95% range, depending on model architecture. Grass and bare land classes show the highest accuracy improvements, from 65.5% and 63.5%, respectively for the RF to the 79.4%-96.3% range for the DLMs. Our work also examined the practical value of having two, instead of one, Landsat sensors. Results indicate substantial classification increases (7%-10% in average F1 accuracy) suggesting that having two concurrent Landsat sensors is important not only for redundancy but also for improved mapping capabilities.