Hi @RyanHemmens , one thing which I would like to highlight is that 20 samples are very less for training a deep learning model because:
One key difference between traditional remote sensing classification approaches and deep learning models is that in traditional methods, such as pixel-based classification, we often rely on a small number of pure pixel samples (e.g., 20-30) to train the model. These methods typically assume that the spectral signature of each class is distinct and sufficient for classification. In contrast, deep learning models, particularly convolutional neural networks (CNNs), do not just learn from the spectral information but also capture higher-level features like shape, texture, and contextual relationships within the image. As a result, deep learning models require larger and more diverse datasets with labeled examples, as they need to learn more complex patterns and relationships that go beyond simple pixel-based information."
You can also try training a DetReg model which can be trained on less samples in comparison to other deep learning object detection models. You can refer to this notebook which shows how DetReg can be used for training a palm tree detection model. Also note that Palm trees have well defined boundaries and their features are different from other trees that is why the model was trained on very small number of samples but Grass trees xanthorrhoea doesn't have well defined boundaries that is why it needs for data.
One more thing which you can try if the above approach doesn't work well is to train a segmentation model instead of objection detection. You can create masks for the Grass trees xanthorrhoea and export the data in Classified Tiles format. Train SamLORA model which can also be trained on less number of samples and achieve accuracy. You can refer this sample notebook for SamLORA workflow: https://developers.arcgis.com/python/latest/samples/finetuning-sam-for-flood-inundation-mapping/
