Investigations on Feature Similarity and the Impact of Training Data for Land Cover Classification

Abstract

Abstract. Fully convolutional neural networks (FCN) are successfully used for pixel-wise land cover classification - the task of identifying the physical material of the Earth’s surface for every pixel in an image. The acquisition of large training datasets is challenging, especially in remote sensing, but necessary for a FCN to perform well. One way to circumvent manual labelling is the usage of existing databases, which usually contain a certain amount of label noise when combined with another data source. As a first part of this work, we investigate the impact of training data on a FCN. We experiment with different amounts of training data, varying w.r.t. the covered area, the available acquisition dates and the amount of label noise. We conclude that the more data is used for training, the better is the generalization performance of the model, and the FCN is able to mitigate the effect of label noise to a high degree. Another challenge is the imbalanced class distribution in most real-world datasets, which can cause the classifier to focus on the majority classes, leading to poor classification performance for minority classes. To tackle this problem, in this paper, we use the cosine similarity loss to force feature vectors of the same class to be close to each other in feature space. Our experiments show that the cosine loss helps to obtain more similar feature vectors, but the similarity of the cluster centers also increases.

Impact of Training Data, Ground Truth and Shape Variability in the Deep Learning-Based Semantic Segmentation of HeLa Cells Observed with Electron Microscopy

“…Similarly, in some cases, the labels of well-defined objects are simplified, i.e., in the context of navigation, moving objects, such as cars or pedestrians, are labeled as moving, and all other objects are labeled as stationary [ 57 ]. In the context of synthetic aperture radar, it is possible to fuse together the image data themselves with a separate source of information, i.e., OpenStreetMap [ 58 ] or extra images from the Earth observation mission Sentinel-2 [ 59 ], to generate labels. In cell biology, and especially as observed with electron microscopy, none of these approaches are feasible due to the inherent variability of cells and other objects, which are less characteristic than a car or a pedestrian and the lack of alternative sources of information at the required resolution.…”

The performance of DST-Wavelet feature extraction for guitar chord recognition

“…This normalization is an attempt to adjust the maximum value of the DWT output signal data to either 1 or -1. The classification algorithms will benefit from this normalization [16]. As a note, the output data from the Normalization 2 process is called feature extraction from the input signal data.…”

Geospatial Information Research: State of the Art, Case Studies and Future Perspectives

“…For label noise, robust methods for training can provide good results even without any handlabeled data, which has been shown for random forest classifiers (Maas et al 2019). These principles are currently being transferred to the domain of deep learning (Voelsen et al 2021).…”