Relating ground photographs to UAV orthomosaics is a key linkage required for accurate multi-scaled lichen mapping. Conventional methods of multi-scaled lichen mapping, such as random forest models and convolutional neural networks, heavily rely on pixel DN values for classification. However, the limited spectral range of ground photos requires additional characteristics to differentiate lichen from spectrally similar objects, such as bright logs. By applying a neural network to tiles of a UAV orthomosaics, additional characteristics, such as surface texture and spatial patterns, can be used for inferences. Our methodology used a neural network (UAV LiCNN) trained on ground photo mosaics to predict lichen in UAV orthomosaic tiles. The UAV LiCNN achieved mean user and producer accuracies of 85.84% and 92.93%, respectively, in the high lichen class across eight different orthomosaics. We compared the known lichen percentages found in 77 vegetation microplots with the predicted lichen percentage calculated from the UAV LiCNN, resulting in a R² relationship of 0.6910. This research shows that AI models trained on ground photographs effectively classify lichen in UAV orthomosaics. Limiting factors include the misclassification of spectrally similar objects to lichen in the RGB bands and dark shadows cast by vegetation.

Lichen mapping is vital for caribou management plans and sustainable land conservation. Previous studies have used random forest, dense neural network, and convolutional neural network models for mapping lichen coverage. However, to date, it is not clear how these models rank in this task. In this study, these machine learning models were evaluated on their ability to predict lichen percent coverage in Sentinel-2 imagery in Québec and Labrador, Canada. The models were trained on 10-m resolution lichen coverage (%) maps created from 20 drone surveys collected in July 2019 and 2022. The dense neural network achieved a higher accuracy than the other two, with a reported mean absolute error of 5.2% and an R² of 0.76. By comparison, the random forest model returned a mean absolute error of 5.5% (R²: 0.74) and the convolutional neural network had a mean absolute error of 5.3% (R²: 0.74). A regional lichen map was created using the trained dense neural network and a Sentinel-2 imagery mosaic. There was greater uncertainty on land covers that the model was not exposed to in training, such as mines and deep lakes. While the dense neural network requires more computational effort to train than a random forest model, the 5.9% performance gain in the test pixel comparison renders it the most suitable for lichen mapping. This study represents progress toward determining the appropriate methodology for generating accurate lichen maps from satellite imagery for caribou conservation and sustainable land management.

Estimating the number of trees within a forest stand, i.e., the forest stand density (FSD), is challenging at large scales. Recently, researchers have turned to a combination of remote sensing and machine learning techniques to derive these estimates. However, in most cases, the developed models rely heavily upon additional data such as LiDAR-based elevations or multispectral information and are mostly applied to managed environments rather than natural/mixed forests. Furthermore, they often require the time-consuming manual digitization or masking of target features, or an annotation using a bounding box rather than a simple point annotation. Here, we introduce the Tree Convolutional Row Neural Network (Tree-CRowNN), an alternative model for tree counting inspired by Multiple-Column Neural Network architecture to estimate the FSD over 12.8 m × 12.8 m plots from high-resolution RGB aerial imagery. Our model predicts the FSD with very high accuracy (MAE: ±2.1 stems/12.8 m², RMSE: 3.0) over a range of forest conditions and shows promise in linking to Sentinel-2 imagery for broad-scale mapping (R²: 0.43, RMSE: 3.9 stems/12.8 m²). We believe that the satellite imagery linkage will be strengthened with future efforts, and transfer learning will enable the Tree-CRowNN model to predict the FSD accurately in other ecozones.

Satellite-derived bathymetry (SDB) is a cost-effective method for estimating water depth in inland and coastal waters, but is only applicable to optically shallow water (OSW). Determining the appropriate extent of SDB maps and the depth threshold for accurate SDB model predictions has therefore been a challenge for practical applications of SDB. Previous studies have used either a numeric cut-off value or manually delineated OSW to determine where to apply, and where not to apply, SDB models. We compared the use of a threshold applied to the predicted depth, automated delineation of OSW using a published model, and manual delineation of OSW, to determine which method of masking unsuitable pixels for SDB performs best. We used a water-leaving reflectance Sentinel-2 image of the St. Lawrence River, and a Random Forest model using neighbouring pixel information to predict SDB. We then compared the different approaches to masking unsuitable pixels in terms of the mean absolute error (MAE) of the retained predictions and the total mapped area. The application of a model-predicted depth threshold is easy to implement and achieved an MAE of 0.54 m, outperforming automated and manual OSW delineation methods, which had MAEs of 1.39 m and 1.64 m respectively over an approximately 100 km² study area. Future studies should further investigate these and other methods for masking pixels unsuitable for SDB under a wider range of environmental conditions.