Browsing by Author "Swaine, Michael"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- ItemImproving the generalisabiility of a deep learning model for global forest classification through image normalisation, enhancement and augmentation(Stellenbosch : Stellenbosch University, 2022-12) Swaine, Michael; Munch, Zahn; Stellenbosch University. Faculty of Arts and Social Sciences. Dept. of Geography and Environmental Studies.ENGLISH ABSTRACT: Effectively managing global forest resources, under threat from climate change, deforestation and fragmentation, requires the efficient extraction of a global tree cover dataset. The purpose of this research was to identify image enhancement and data augmentation methods that would improve the generalisability of a deep learning model for the classification of global tree cover. In the first experiment we aimed to improve the accuracy of a deep learning model for global forest classification using Sentinel 2 optical data. We present several image enhancement methods widely used in natural image classification and biomedical imaging domains, including histogram equalisation (HE), contrast limited adaptive histogram equalisation (CLAHE) and global contrast normalisation (GCN), as pre-processing steps. The enhancement methods were compared with each other on a per biome basis, and both training and validation regions were selected to represent the heterogeneity within biomes. Selected images were captured within the local optimal foliage growing season and contained minimal or no clouds. A U-Net convolutional neural network model was trained for each enhancement per biome and used to perform inference on validation images for each of the corresponding biomes and enhancements. Random stratified samples were collated for all validation images per biome per enhancement for statistical analysis. Only GCN and CLAHE RGB returned higher means than the baseline dataset. The results showed that GCN most consistently improved classification results for tree cover across biomes, possibly due to the standardization of contrast levels of the training and validation images. In the absence of accurately annotated training data for tree segmentation, training a robust, deep learning model for global tree cover classification remains a challenge. As its first objective, experiment 2 evaluated basic data augmentation methods and prediction frameworks that might lead to achieving an accurate, global tree cover classification. A training dataset was artificially inflated using common geometric and colour data augmentation methods borrowed from the computer vision domain. Their effectiveness in improving the generalisability of a U-Net model for tree classification was tested. Both geometric and colour augmentations, when applied individually, showed improvements in model accuracy. When applied together, the combined augmentations showed only marginal improvements over the individually applied augmentations. The second objective was to test two approaches towards achieving a global tree classification. The first was a model per biome approach, whereby a model was trained with data derived only from the respective biome. The second involved training a single globally representative model with training data from all biomes combined. This resulted in higher MCC scores than the multi-model approach. The diversity in training data appeared to increase model robustness. Thus, it was found that training a single, globally representative model with a combination of colour and geometric augmentations led to an effective framework to infer a global tree classification