Wetland Classification with Deep ResU-Net Convolutional Neural Network and Multitemporal Sentinel-1 & 2 Imagery and ALOS Elevation Data: A Case Study in Alberta Parkland & Grassland Natural Region, Canada

The study aimed to develop a Deep Learning (DL) model for a large-scale wetland classification in Alberta's Parkland and Grassland Natural Region (PGNR) using a fusion of multi-temporal Sentinel-2 (S2) optical and Sentinel-1 (S1) radar data and topographic data. A key objective of the study was to compare the performance of the ResNet model with two shallow learning techniques (namely Random Forest (RF) and Support Vector Machine (SVM)). A 25-band multi-seasonal (acquired over the summer/fall months of 2017 to 2020) image stack comprised of S1 (dual-polarization vertical-horizontal (VH) bands) and S2 (near-infrared (band 8) and shortwave infrared (band 11)) images and Advanced Land Observing Satellite (ALOS) derived Topographic Wetness Index as input data in the three models. Comparing the three products' accuracy metrics showed that the CNN model significantly outperformed the shallow machine learning models (SVM and RF). The best performing model was the ResU-Net model, with overall accuracy and overall kappa of 74% and 0.62, followed by SVM (69% and 0.55) and RF (0.68 and 0.54), respectively. The relative F1 scores of the mapped wetlands (marsh, open water, and swamp) using the shallow ML models showed deficiencies in their predictive capabilities. The average F1 score of the ResNet model was 0.77 compared to 0.65 (for SVM) and 0.64 (for RF). Compared to the ResNet CNN predictions, it was evident that this DL technique outperformed the shallow ML techniques evaluated in the study.

本研究旨在开发深度学习（Deep Learning, DL）模型，以对加拿大阿尔伯塔省草原与公园自然区（Parkland and Grassland Natural Region, PGNR）开展大规模湿地分类，研究融合了多时序哨兵-2（Sentinel-2, S2）光学影像、哨兵-1（Sentinel-1, S1）雷达数据与地形数据。本研究的核心目标之一是对比ResNet模型与两种浅层学习方法——随机森林（Random Forest, RF）和支持向量机（Support Vector Machine, SVM）的分类性能。研究所采用的输入数据为包含25个波段的多季相影像集，采集时段为2017至2020年的夏秋两季，该影像集涵盖S1影像（垂直-水平双极化（VH）波段）、S2影像（近红外波段（第8波段）与短波红外波段（第11波段）），以及先进陆地观测卫星（Advanced Land Observing Satellite, ALOS）衍生的地形湿度指数。对三类模型的精度指标进行对比后发现，卷积神经网络（Convolutional Neural Network, CNN）模型的性能显著优于浅层机器学习模型（SVM与RF）。其中表现最优的模型为ResU-Net，其总体精度与总体Kappa系数分别为74%与0.62；紧随其后的是SVM（总体精度69%、总体Kappa系数0.55），RF则以总体精度68%、总体Kappa系数0.54位列第三。利用浅层机器学习模型得到的湿地分类结果（包括草本沼泽、开阔水体与木本沼泽）的相对F1分数显示，此类模型的预测能力存在明显不足。ResNet模型的平均F1分数为0.77，而SVM与RF的平均F1分数分别为0.65与0.64。相较于本研究中评估的ResNet卷积神经网络预测结果，深度学习技术的整体表现显著优于浅层机器学习方法。