Random Forest (RF) is a widely used algorithm for classification of remotely sensed data. Through a case study in peatland classification using LiDAR derivatives, we present an analysis of the effects of input data characteristics on RF classifications (including RF out-of-bag error, independent classification accuracy and class proportion error). Training data selection and specific input variables (i.e., image channels) have a large impact on the overall accuracy of the image classification. High-dimension datasets should be reduced so that only uncorrelated important variables are used in classifications. Despite the fact that RF is an ensemble approach, independent error assessments should be used to evaluate RF results, and iterative classifications are recommended to assess the stability of predicted classes. Results are also shown to be highly sensitive to the size of the training data set. In addition to being as large as possible, the training data sets used in RF classification should also be (a) randomly distributed or created in a manner that allows for the class proportions of the training data to be representative of actual class proportions in the landscape; and (b) should have minimal spatial autocorrelation to improve classification results and to mitigate inflated estimates of RF out-of-bag classification accuracy.

Random Forests variable importance measures are often used to rank variables by their relevance to a classification problem and subsequently reduce the number of model inputs in high-dimensional data sets, thus increasing computational efficiency. However, as a result of the way that training data and predictor variables are randomly selected for use in constructing each tree and splitting each node, it is also well known that if too few trees are generated, variable importance rankings tend to differ between model runs. In this letter, we characterize the effect of the number of trees (ntree) and class separability on the stability of variable importance rankings and develop a systematic approach to define the number of model runs and/or trees required to achieve stability in variable importance measures. Results demonstrate that both a large ntree for a single model run, or averaged values across multiple model runs with fewer trees, are sufficient for achieving stable mean importance values. While the latter is far more computationally efficient, both the methods tend to lead to the same ranking of variables. Moreover, the optimal number of model runs differs depending on the separability of classes. Recommendations are made to users regarding how to determine the number of model runs and/or trees that are required to achieve stable variable importance rankings.

In this paper, we assess the use of Random Forest (RF) for mapping land cover classes within Mer Bleue bog, a large northern peatland in southeastern Ontario near Ottawa, Canada, using Synthetic Aperture Radar (SAR) and airborne Light Detection and Ranging (LiDAR). Not only has RF been shown to improve classification accuracies over more traditional classifiers, but it also provides useful information on the statistical importance of individual input image bands for land cover classification. Our specific objectives in this study were to: (i) assess the robustness of a RF approach to northern peatland classification; (ii) examine variable importance resulting from the RF classifications to identify which imagery types, derivatives, and analysis scales are most useful for mapping different classes of northern peatlands; (iii) assess if fusion of different LiDAR and SAR variables can improve classification accuracies at Mer Bleue; and (iv) assess physical interpretability of the multisensor image types and ... Dans cet article, on évalue l'utilisation de la technique des forêts aléatoires («Random Forest», RF) pour la cartographie des classes de couvert dans la tourbière de la Mer Bleue, une vaste tourbière nordique située dans le sud-est de l'Ontario près d'Ottawa, au Canada, à l'aide de données RSO (radar à synthèse d'ouverture) et LiDAR («Light Detection and Ranging») aéroporté. Non seulement il a été démontré que la méthode RF améliorait la précision de classification par rapport aux classifieurs plus traditionnels, mais elle apporte également une information utile sur l'importance statistique des bandes individuelles de l'image d'entrée pour la classification du couvert. Nos objectifs précis dans cette étude étaient : (i) d’évaluer la robustesse d'une approche RF pour la classification des tourbières nordiques; (ii) d'examiner l'importance des variables résultant des classifications RF pour identifier quels types d'image, dérivées et échelles d'analyse sont les plus utiles pour cartographier les différente...

The following review is the second part of a two part series on the use of remotely sensed data for quantifying wetland extent and inferring or measuring condition for monitoring drivers of change on wetland environments. In the first part, we introduce policy makers and non-users of remotely sensed data with an effective feasibility guide on how data can be used. In the current review, we explore the more technical aspects of remotely sensed data processing and analysis using case studies within the literature. Here we describe: (a) current technologies used for wetland assessment and monitoring; (b) the latest algorithmic developments for wetland assessment; (c) new technologies; and (d) a framework for wetland sampling in support of remotely sensed data collection. Results illustrate that high or fine spatial resolution pixels (≤10 m) are critical for identifying wetland boundaries and extent, and wetland class, form and type, but are not required for all wetland sizes. Average accuracies can be up to 11% better (on average) than medium resolution (11–30 m) data pixels when compared with field validation. Wetland size is also a critical factor such that large wetlands may be almost as accurately classified using medium-resolution data (average = 76% accuracy, stdev = 21%). Decision-tree and machine learning algorithms provide the most accurate wetland classification methods currently available, however, these also require sampling of all permutations of variability. Hydroperiod accuracy, which is dependent on instantaneous water extent for single time period datasets does not vary greatly with pixel resolution when compared with field data (average = 87%, 86%) for high and medium resolution pixels, respectively. The results of this review provide users with a guideline for optimal use of remotely sensed data and suggested field methods for boreal and global wetland studies.

Despite the enormous success of the ’pre-training and fine-tuning’ paradigm, widespread across machine learning, it has yet to pervade remote sensing (RS). To help rectify this, we pre-train a vision transformer (ViT) on 1.3 million satellite-derived RS images. We pre-train SatViT using a state-of-the-art self-supervised learning algorithm called masked autoencoding (MAE), which learns general representations by reconstructing held-out image patches. Crucially, this approach does not require annotated data, allowing us to pre-train on unlabeled images acquired from Sentinel-1 & 2. After fine-tuning, SatViT outperforms state-of-the-art ImageNet and RS-specific pre-trained models on both of our downstream tasks. We further improve overall accuracy (by 3.2% and 0.21%) by continuing to pre-train SatViT—still using MAE—on the unlabelled target datasets. Most importantly, we release our code, pre-trained model weights, and tutorials aimed at helping researchers fine-tune our models. (https://github.com/antofuller/SatViT).