The Python Landscape Classifier (PyLC): A new deep learning semantic segmentation workflow for land cover classification of oblique ground-based mountain photographs

Automated land cover classification is a common practice for quantifying environmental change at large scales. Most progress to date has focused on classifying nadir imagery. This ignores a vast source of land cover data in the form of oblique images (i.e., at an angle not perpendicular to the ground). Recently, software has emerged for converting oblique images into georeferenced forms, opening the door to quantitative analysis, yet the lack of automated methods for classifying land cover in oblique images remains a major roadblock. Here we present the Python Landscape Classifier (PyLC), a Pytorch-based trainable segmentation workflow using DeeplabV3+ for automated classification of land cover in oblique images. Pre-trained models for PyLC were developed using the MountainScape Segmentation Dataset comprising manually segmented high-resolution images of Canadian mountain landscapes. Models exist for both historical (grayscale) images and modern (color) images with F1 scores of 0.60 and 0.77 and weighted IoU of 0.66 and 0.78, respectively. We compared the base performance of PyLC with DeeplabV3+ to three other model architectures specifically designed for ultra-high-resolution images. DeeplabV3+ outperformed all other architectures except when classifying historical images where ISDNet surpassed DeeplabV3+ in terms of F1 score and weighted IoU by 1% and 4%, respectively.