Search for: All records

Semantic segmentation algorithms, such as UNet, that rely on convolutional neural network (CNN)-based architectures, due to their ability to capture local textures and spatial context, have shown promise for anthropogenic geomorphic feature extraction when using land surface parameters (LSPs) derived from digital terrain models (DTMs) as input predictor variables. However, the operationalization of these supervised classification methods is limited by a lack of large volumes of quality training data. This study explores the use of transfer learning, where information learned from another, and often much larger, dataset is used to potentially reduce the need for a large, problem-specific training dataset. Two anthropogenic geomorphic feature extraction problems are explored: the extraction of agricultural terraces and the mapping of surface coal mine reclamation-related valley fill faces. Light detection and ranging (LiDAR)-derived DTMs were used to generate LSPs. We developed custom transfer parameters by attempting to predict geomorphon-based landforms using a large dataset of digital terrain data provided by the United States Geological Survey’s 3D Elevation Program (3DEP). We also explored the use of pre-trained ImageNet parameters and initializing models using parameters learned from the other mapping task investigated. The geomorphon-based transfer learning resulted in the poorest performance while the ImageNet-based parameters generally improved performance in comparison to a random parameter initialization, even when the encoder was frozen or not trained. Transfer learning between the different geomorphic datasets offered minimal benefits. We suggest that pre-trained models developed using large, image-based datasets may be of value for anthropogenic geomorphic feature extraction from LSPs even given the data and task disparities. More specifically, ImageNet-based parameters should be considered as an initialization state for the encoder component of semantic segmentation architectures applied to anthropogenic geomorphic feature extraction even when using non-RGB image-based predictor variables, such as LSPs. The value of transfer learning between the different geomorphic mapping tasks may have been limited due to smaller sample sizes, which highlights the need for continued research in using unsupervised and semi-supervised learning methods, especially given the large volume of digital terrain data available, despite the lack of associated labels. 

Convolutional neural network (CNN)-based deep learning (DL) methods have transformed the analysis of geospatial, Earth observation, and geophysical data due to their ability to model spatial context information at multiple scales. Such methods are especially applicable to pixel-level classification or semantic segmentation tasks. A variety of R packages have been developed for processing and analyzing geospatial data. However, there are currently no packages available for implementing geospatial DL in the R language and data science environment. This paper introduces the geodl R package, which supports pixel-level classification applied to a wide range of geospatial or Earth science data that can be represented as multidimensional arrays where each channel or band holds a predictor variable. geodl is built on the torch package, which supports the implementation of DL using the R and C++ languages without the need for installing a Python/PyTorch environment. This greatly simplifies the software environment needed to implement DL in R. Using geodl, geospatial raster-based data with varying numbers of bands, spatial resolutions, and coordinate reference systems are read and processed using the terra package, which makes use of C++ and allows for processing raster grids that are too large to fit into memory. Training loops are implemented with the luz package. The geodl package provides utility functions for creating raster masks or labels from vector-based geospatial data and image chips and associated masks from larger files and extents. It also defines a torch dataset subclass for geospatial data for use with torch dataloaders. UNet-based models are provided with a variety of optional ancillary modules or modifications. Common assessment metrics (i.e., overall accuracy, class-level recalls or producer’s accuracies, class-level precisions or user’s accuracies, and class-level F1-scores) are implemented along with a modified version of the unified focal loss framework, which allows for defining a variety of loss metrics using one consistent implementation and set of hyperparameters. Users can assess models using standard geospatial and remote sensing metrics and methods and use trained models to predict to large spatial extents. This paper introduces the geodl workflow, design philosophy, and goals for future development. 

Evaluating classification accuracy is a key component of the training and validation stages of thematic map production, and the choice of metric has profound implications for both the success of the training process and the reliability of the final accuracy assessment. We explore key considerations in selecting and interpreting loss and assessment metrics in the context of data imbalance, which arises when the classes have unequal proportions within the dataset or landscape being mapped. The challenges involved in calculating single, integrated measures that summarize classification success, especially for datasets with considerable data imbalance, have led to much confusion in the literature. This confusion arises from a range of issues, including a lack of clarity over the redundancy of some accuracy measures, the importance of calculating final accuracy from population-based statistics, the effects of class imbalance on accuracy statistics, and the differing roles of accuracy measures when used for training and final evaluation. In order to characterize classification success at the class level, users typically generate averages from the class-based measures. These averages are sometimes generated at the macro-level, by taking averages of the individual-class statistics, or at the micro-level, by aggregating values within a confusion matrix, and then, calculating the statistic. We show that the micro-averaged producer’s accuracy (recall), user’s accuracy (precision), and F1-score, as well as weighted macro-averaged statistics where the class prevalences are used as weights, are all equivalent to each other and to the overall accuracy, and thus, are redundant and should be avoided. Our experiment, using a variety of loss metrics for training, suggests that the choice of loss metric is not as complex as it might appear to be, despite the range of choices available, which include cross-entropy (CE), weighted CE, and micro- and macro-Dice. The highest, or close to highest, accuracies in our experiments were obtained by using CE loss for models trained with balanced data, and for models trained with imbalanced data, the highest accuracies were obtained by using weighted CE loss. We recommend that, since weighted CE loss used with balanced training is equivalent to CE, weighted CE loss is a good all-round choice. Although Dice loss is commonly suggested as an alternative to CE loss when classes are imbalanced, micro-averaged Dice is similar to overall accuracy, and thus, is particularly poor for training with imbalanced data. Furthermore, although macro-Dice resulted in models with high accuracy when the training used balanced data, when the training used imbalanced data, the accuracies were lower than for weighted CE. In summary, the significance of this paper lies in its provision of readers with an overview of accuracy and loss metric terminology, insight regarding the redundancy of some measures, and guidance regarding best practices.