A novel weakly supervised semantic segmentation framework to improve the resolution of land cover product

A Weakly Supervised Semantic Segmentation Framework for Medium-Resolution Forest Classification With Noisy Labels and GF-1 WFV Images

“…In recent years, deep learning has demonstrated great potential in feature extraction and has been applied by scholars to remote sensing information extraction [44][45][46]. It has exhibited superior performance across various objects [32,47], such as impervious surfaces/built-up areas, crop, water and vegetation [30,48]. Deep learning algorithms rely heavily on massive fine-labeled samples.…”

“…At present, there many studies applying weakly supervised techniques and deep learning algorithms to the classification task of medium-resolution remote sensing images [30]. Several studies have applied weakly supervised deep learning methods in LULC classification tasks [32,33] with promising results, such as paddy rice mapping [59]. Moreover, some scholars have reported some global-or national-scale LULC products based on deep learning algorithms, such as Esri land cover and CRLC [39].…”

“…To eliminate the influence of resolution mismatch and semantic errors in the labels, Liu et al [39] modeled the spectral similarity and spatial adjacency of the training labels using a conditional random field to refine the training labels and used a class-conditional label correction methods to detect and correct the abnormal incorrect labels. To promote the accuracy of the labels, Chen et al [32] trained a weakly supervised learning network using partially labeled hyper-pixels generated from the representative samples which are transferred from low-resolution LULC products; Zhang et al [67] performed pre-classification based on a limited number of point samples and a support vector machine algorithm and took the high-confidence areas as training samples for the deep learning model; Liu et al [41] create the sample by integrated existing LULC products and OSM. > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <…”

“…Moreover, such algorithms rely heavily on massive finely labeled samples which are challenging to acquire. The recent development of weakly supervised technology [31][32][33] reduces the demand for sample labeling, demonstrating the robustness of the models to inaccurate, incomplete and inexact labels [34]. This suggests that the deep learning methods are promising for forest research [30,35].…”

“…Although such an approach might lead to errors in the training data, a certain percentage of errors barely affects the classification results [9,14]. There are some methods that are usually used to reduce the labeling error to improve the accuracy of the samples [17,32], such as voting [41] and D-S evidence theory [42]. A few recent studies have successfully realized the large-scale LULC classification with existing LULC products as labels [39,40,43].…”