Land Cover Classification Accuracy Research Articles

Semantic Segmentation of Hyperspectral Remote Sensing Images Based on PSE-UNet Model.

With the development of deep learning, the use of convolutional neural networks (CNN) to improve the land cover classification accuracy of hyperspectral remote sensing images (HSRSI) has become a research hotspot. In HSRSI semantics segmentation, the traditional dataset partition method may cause information leakage, which poses challenges for a fair comparison between models. The performance of the model based on "convolutional-pooling-fully connected" structure is limited by small sample sizes and high dimensions of HSRSI. Moreover, most current studies did not involve how to choose the number of principal components with the application of the principal component analysis (PCA) to reduce dimensionality. To overcome the above challenges, firstly, the non-overlapping sliding window strategy combined with the judgment mechanism is introduced, used to split the hyperspectral dataset. Then, a PSE-UNet model for HSRSI semantic segmentation is designed by combining PCA, the attention mechanism, and UNet, and the factors affecting the performance of PSE-UNet are analyzed. Finally, the cumulative variance contribution rate (CVCR) is introduced as a dimensionality reduction metric of PCA to study the Hughes phenomenon. The experimental results with the Salinas dataset show that the PSE-UNet is superior to other semantic segmentation algorithms and the results can provide a reference for HSRSI semantic segmentation.

AMFFNet: attention-guided multi-level feature fusion network for land cover classification of remote sensing images

In the field of remote sensing, the classification of land cover is a pivotal and challenging issue. Standard models fail to capture global and semantic information in remote sensing images despite the fact that a convolutional neural network provides robust support for semantic segmentation. In addition, owing to disparities in semantic levels and spatial resolution, the simple fusion of low-level and high-level features may diminish the efficiency. To address these deficiencies, an attention-guided multi-level feature fusion network (AMFFNet) is proposed in this study. The proposed AMFFNet approach is designed as an encoder–decoder network with the inclusion of a multi-level feature fusion module (MFF) and a dual attention map module (DAM). A DAM models the semantic association of features from a spatial and channel perspective, and an MFF bridges the semantic and resolution gaps between high-level and low-level features. Furthermore, we propose a residual-based boundary refinement upsample module to further optimize the object boundaries. The experimental results indicate that the proposed strategy can considerably enhance the accuracy of land cover classification, achieving a mean intersection over union of 90.39% on the LandCover.ai dataset and 63.14% on the Gaofen Image Dataset with 15 categories (GID-15).

Mapping the spatial and temporal patterns of fallow land in mountainous regions of China

ABSTRACT The rapid growth of the global population has resulted in a continuous increase in cropland intensity and a shortening of the fallow period as part of the cropland rotation cycle. Yet, there is a lack of systematic knowledge on the extent of fallow lands, particularly in complex landscapes, such as the mountainous regions of China. To fill this knowledge gap, taking Yuanyang County (YYC), Yunnan Province, China, as a case study, we tested a method to identify the spatial-temporal distribution of fallow land by mapping cropland with Landsat data. The overall accuracy of land cover classification, including cropland, ranged between 90.1% and 95.8% from 1998 to 2019. The average accuracy of fallow plots was 75.7% from 2001 to 2019. The annual fallow rate varied between 8.3% and 54.3%, with an average of 20.7%. Kernel density estimated with the probability density function showed that fallow varied between 5 and 13 blocks per km2, gradually decreasing from the central area to the periphery. Increasing elevation, the low value of regional domestic products, and the increased distance to rural settlements were closely related to the higher proportions of fallow land. The approach presented here can be applied to map fallow land in other regions.

An MLC and U-Net Integrated Method for Land Use/Land Cover Change Detection Based on Time Series NDVI-Composed Image from PlanetScope Satellite

Land use/land cover change (LUCC) detection based on optical remote-sensing images is an important research direction in the field of remote sensing. The key to it is to select an appropriate data source and detection method. In recent years, the continuous expansion of construction land in urban areas has become the main reason for the increase in LUCC demand. However, due to the complexity and diversity of land-cover types, it is difficult to obtain high-precision classification results. In this article, a 12-month time series NDVI (Normalized Difference Vegetation Index) image of the study area was generated based on the high spatial and temporal resolution PlanetScope satellite images. According to the time series NDVI image, representative land-cover samples were selected, and the changed land samples were selected at the same time. This method could directly obtain the LUCC detection results of the study area through land-cover classification. First, Maximum Likelihood Classification (MLC), a classical machine-learning method, was used for supervised classification, and the samples needed for deep learning were selected according to the classification results. Then, the U-Net model, which can fully identify and explore the deep semantic information of the time series NDVI image, was used for land classification. Finally, this article made a comparative analysis of the two classification results. The results demonstrate that the overall classification accuracy based on time series NDVI is significantly higher than that of single-scene NDVI and mean NDVI. The LUCC detection method proposed in this article can effectively extract changed areas. The overall accuracy of the MLC and U-Net model is 79.38% and 85.26%, respectively. Therefore, the deep-learning method can effectively improve the accuracy of land-cover classification and change detection.

Hyperspectral image classification based on combining texture and multiscale guided filter feature of principal component

Combining spectral and spatial features can increase the land-cover classification accuracy of hyperspectral images (HSI), but improvements are still needed to overcome problems remain in the extraction and use of spatial features, such as fixed extraction scale, loss of homogeneous region boundary and texture detail. Therefore, a novel classification frame is constructed herein. First, the first principal component of a Karhunen–Loeve transform of the HSI is taken as the guidance image to extract guide filter features of the other selected principal components under different scales and the appropriate multiscale range is automatically determined by calculating the interclass distance and intraclass distance based on the training samples. Second, the texture features, extracted by a local binary pattern, are used to emphasize the spatial information. Finally, logistic regression via variable splitting and augmented Lagrangian is used to obtain class labels from each scale image feature, and the majority voting algorithm is used to obtain the last class label of each pixel. Two standard HSI datasets, Indian Pines and Pavia University, are used to test this frame, and methods based on spectrum or spatial filtering (Type I) and methods combining spectral and spatial features (Type II) are used for comparison. The overall accuracy (OA) of the proposed frame is on average 13.76% and 4.77% higher than that of Type I and Type II, respectively, when 30 pixels per class are used as training samples. One dataset with a spatial resolution of 0.8 m, obtained by Panchromatic multispectral imaging spectrometer, is used for comparison with Type I, resulting in an average increase in OA of 7.05%, when 30 pixels per class are used as training samples. These results indicate the validity of the proposed frame.

Dynamic Convolution Self-Attention Network for Land-Cover Classification in VHR Remote-Sensing Images

The current deep convolutional neural networks for very-high-resolution (VHR) remote-sensing image land-cover classification often suffer from two challenges. First, the feature maps extracted by network encoders based on vanilla convolution usually contain a lot of redundant information, which easily causes misclassification of land cover. Moreover, these encoders usually require a large number of parameters and high computational costs. Second, as remote-sensing images are complex and contain many objects with large-scale variances, it is difficult to use the popular feature fusion modules to improve the representation ability of networks. To address the above issues, we propose a dynamic convolution self-attention network (DCSA-Net) for VHR remote-sensing image land-cover classification. The proposed network has two advantages. On one hand, we designed a lightweight dynamic convolution module (LDCM) by using dynamic convolution and a self-attention mechanism. This module can extract more useful image features than vanilla convolution, avoiding the negative effect of useless feature maps on land-cover classification. On the other hand, we designed a context information aggregation module (CIAM) with a ladder structure to enlarge the receptive field. This module can aggregate multi-scale contexture information from feature maps with different resolutions using a dense connection. Experiment results show that the proposed DCSA-Net is superior to state-of-the-art networks due to higher accuracy of land-cover classification, fewer parameters, and lower computational cost. The source code is made public available.

UAV-Based Multi-Sensor Data Fusion for Urban Land Cover Mapping Using a Deep Convolutional Neural Network

Accurate and up-to-date land cover classification information is essential for many applications, such as land-use change detection, global environmental change, and forest management, among others. Unoccupied aerial systems (UASs) provide the advantage of flexible and rapid data acquisition at low cost compared to conventional platforms, such as satellite and airborne systems. UASs are often equipped with high spatial resolution cameras and/or light detection and ranging (LiDAR). However, the high spatial resolution imagery has a high information content, which makes land cover classification quite challenging. Recently, deep convolutional neural networks (DCNNs) have been effectively applied to remote sensing applications, which overcome the drawback of traditional techniques. In this research, a low-cost UAV-based multi-sensor data fusion model was developed for land cover classification based on a DCNN. For the purpose of this research, two datasets were collected at two different urban locations using two different UASs. A DCNN model, based on U-net with Resnet101 as a backbone, was used to train and test the fused image/LiDAR data. The maximum likelihood and support vector machine techniques were used as a reference for classifier comparison. It was shown that the proposed DCNN approach improved the overall accuracy of land cover classification for the first dataset by 15% compared to the reference classifiers. In addition, the overall accuracy of land cover classification improved by 7%, and the precision, recall, and F-measure improved by 18% when the fused image/LiDAR data were used compared to the images only. The trained DCNN model was also tested on the second dataset, and the obtained results were largely similar to those of the first dataset.

Learning latent local manifold embeddings for hyperspectral feature extraction

Hyperspectral images with an immense number of spectral bands provide abundant discriminant information for accurate land-cover classification in the remote sensing field. However, these narrow and adjacent bands contain a large amount of redundant information. Analyzing these images always requires a huge storage space with expensive computational costs. Furthermore, their high correlation coefficient would lead to the Hughes phenomenon, hindering the improvement of classification performance. We propose a linear semi-supervised hyperspectral feature extraction method L3ME to learn latent local manifold embeddings. Although labeled samples are beneficial to construct learning models, their number is always limited in real-world tasks. The motivation of this paper is to jointly enhance the contributions of labeled and unlabeled samples for learning local manifold structures of hyperspectral images. Features in labeled samples are extracted by two procedures, the adaptive patch alignment framework and integrated intraclass-interclass relationships, from different perspectives. The former aims to solve the problem of the uneven distribution of classes by introducing spectral angle based adaptive parameters. The latter aims to solve the problem of the uneven distribution of samples by constructing several adjacency graphs. The locality preserving projection is capable of preserving the local neighborhood structure of samples. A penalty for sparse regularization is cleverly integrated into the proposed linear discriminant objective function, which is optimized using a novel updating strategy. The convergence of L3ME is proved in detail and analyzed in this paper. Experiments on three typical hyperspectral datasets illustrate the effectiveness of the proposed method over some state-of-the-art techniques. The implementation of L3ME is available at https://github.com/biowby/L3ME.

HFENet: Hierarchical Feature Extraction Network for Accurate Landcover Classification

Landcover classification is an important application in remote sensing, but it is always a challenge to distinguish different features with similar characteristics or large-scale differences. Some deep learning networks, such as UperNet, PSPNet, and DANet, use pyramid pooling and attention mechanisms to improve their abilities in multi-scale features extraction. However, due to the neglect of low-level features contained in the underlying network and the information differences between feature maps, it is difficult to identify small-scale objects. Thus, we propose a novel image segmentation network, named HFENet, for mining multi-level semantic information. Like the UperNet, HFENet adopts a top-down horizontal connection architecture while includes two improved modules, the HFE and the MFF. According to the characteristics of different levels of semantic information, HFE module reconstructs the feature extraction part by introducing an attention mechanism and pyramid pooling module to fully mine semantic information. With the help of a channel attention mechanism, MFF module up-samples and re-weights the feature maps to fuse them and enhance the expression ability of multi-scale features. Ablation studies and comparative experiments between HFENet and seven state-of-the-art models (U-Net, DeepLabv3+, PSPNet, FCN, UperNet, DANet and SegNet) are conducted with a self-labeled GF-2 remote sensing image dataset (MZData) and two open datasets landcover.ai and WHU building dataset. The results show that HFENet on three datasets with six evaluation metrics (mIoU, FWIoU, PA, mP, mRecall and mF1) are better than the other models and the mIoU is improved 7.41–10.60% on MZData, 1.17–11.57% on WHU building dataset and 0.93–4.31% on landcover.ai. HFENet can perform better in the task of refining the semantic segmentation of remote sensing images.

KAJIAN PEMANTAUAN PERUBAHAN PENUTUPAN LAHAN BERBASIS PENGINDERAAN JAUH DI KOTAWARINGIN BARAT

The first step in monitoring deforestation is to measure and identify land cover changes. Our research uses remote sensing technology, because it has advantages such as: cheap, transparent, fast and has measurable accuracy and temporal analysis. This study aims to identify, measure, and analyze trends in land cover change in 1990-2000, 2000-2012, and 2012-2021. The results showed that there were 18 land cover classes with details of 3 forest classes; and 15 non-forest classes; Forest class decreased throughout 1990 to 2021, the trend of changing forest cover was deforestation into plantations, swamp shrubs, dryland agriculture mixed with shrubs, and swamps. The rate of deforestation in 1990-2000 was 3.07%, deforestation in 2000-2012 was 3.02%, and deforestation in 2012-2021 was 1.02%. The accuracy of land cover classification is 91%.

Enhancing the Accuracy of Land Cover Classification by Airborne LiDAR Data and WorldView-2 Satellite Imagery

The Full Waveform LiDAR system has been developed and used commercially all over the world. It acts to record the complete time of a laser pulse and has a high-resolution sampling interval compared to the traditional multiple-echo LiDAR, which only provides signals within a single target range. This study area mainly collects data from Riegl LMS-Q680i Full Waveform LiDAR and WorldView-2 satellite imagery, which focuses on buildings, vegetation, grassland, asphalt roads and other ground types as the surface objects. The amplitude and pulse width are selected as waveform basic parameters. The parameter of topography is slope, and the height classification parameters of the test ground are 0–0.5 m, 0.5–2.5 m, and 2.5 m. To eliminate noise, the neighborhood average is applied on the LiDAR parameter values and analyzed as the classification accuracy comparison. This survey uses Decision Tree as the classification method. Comparing the data between neighborhood average and non-neighborhood average, the data classification accuracy improves by 7%, and Kappa improves by 5.92%. NDVI image data are utilized to distinguish the artificial from natural ground. The results show that the neighborhood average with previous data can improve the classification accuracy by 5%, and Kappa improves by 4.25%. By adding NIR-2 of WorldView-2 satellite imagery to the neighborhood average analysis, the overall classification accuracy is improved by 2%, and the Kappa value by 1.21%. This article shows that utilizing the analysis of neighborhood average and image parameters can effectively improve the classification accuracy of land covers.

CapViT: Cross-context capsule vision transformers for land cover classification with airborne multispectral LiDAR data

Equipped with multiple channels of laser scanners, multispectral light detection and ranging (MS-LiDAR) devices possess more advanced prospects in earth observation tasks compared with their single-band counterparts. It also opens up a potential-competitive solution to conducting land cover mapping with MS-LiDAR devices. In this paper, we develop a cross-context capsule vision transformer (CapViT) to serve for land cover classification with MS-LiDAR data. Specifically, the CapViT is structurized with three streams of capsule transformer encoders, which are stacked by capsule transformer (CapFormer) blocks, to exploit long-range global feature interactions at different context scales. These cross-context feature semantics are finally effectively fused to supervise accurate land cover type inferences. In addition, the CapFormer block parallels dual-path multi-head self-attention modules functioning to interpret both spatial token correlations and channel feature interdependencies, which favor significantly to the semantic promotion of feature encodings. Consequently, with the semantic-promoted feature encodings to boost the feature representation distinctiveness and quality, the land cover classification accuracy is effectively improved. The CapViT is elaborately testified on two MS-LiDAR datasets. Both quantitative assessments and comparative analyses demonstrate the competitive capability and advanced performance of the CapViT in tackling land cover classification issues.

Evaluation of SAR and Optical Image Fusion Methods in Oil Palm Crop Cover Classification Using the Random Forest Algorithm

This paper presents an evaluation of land cover accuracy, particularly regarding oil palm crop cover, using optical/synthetic aperture radar (SAR) image fusion methods through the implementation of the random forest (RF) algorithm on cloud computing platforms using Sentinel-1 SAR and Sentinel-2 optical images. Among the fusion methods evaluated were Brovey (BR), high-frequency modulation (HFM), Gram–Schmidt (GS), and principal components (PC). This work was developed using a cloud computing environment employing R and Python for statistical analysis. It was found that an optical/SAR image stack resulted in the best overall accuracy with 82.14%, which was 11.66% higher than that of the SAR image, and 7.85% higher than that of the optical image. The high-frequency modulation (HFM) and Brovey (BR) image fusion methods showed overall accuracies higher than the Sentinel-2 optical image classification by 3.8% and 3.09%, respectively. This demonstrates the potential of integrating optical imagery with Sentinel SAR imagery to increase land cover classification accuracy. On the other hand, the SAR images obtained very high accuracy results in classifying oil palm crops and forests, reaching 94.29% and 90%, respectively. This demonstrates the ability of synthetic aperture radar (SAR) to provide more information when fused with an optical image to improve land cover classification.

DEEP LEARNING TRAINING WITH UNBALANCE SAMPLE DISTRIBUTION FOR REMOTE SENSING IMAGE SEGMENTATION

Abstract. The intelligent interpretation of remote sensing images based on deep learning has become a hot spot with the increasing satellite images acquired due to the rapid development of aerospace technology. Sufficient and reasonable distributed samples are essential for the accuracy of deep learning. The spatial distribution of natural features is inhomogeneous in the real world. When people create sample dataset, they often collect within a certain local range, which may bring problems of unbalanced distribution of samples, including the unbalance between training dataset and validation dataset, and the unbalance among different sample categories. This long-tail distribution of samples (i.e., a few classes account for most of the data, while most classes are under-represented) can lead to bias in the training model and make it difficult to ensure accuracy.In this paper we tried to solved the above-mentioned problem in landcover classification with high spatial and spectral resolution (HSSR) remote sensing images. We first adopted an iterative stratification method for multi-label data classification to ensure that both training dataset and validation dataset contain reasonable proportion of landcover classes. Then we proposed a weighted loss algorithm to further strengthen the learning ability of the model for rare categories. Experiments on a large volume HSSR dataset shows that with our methods the accuracy of landcover classification increased by 2%.

Monitoring urbanization and environmental impact in Kigali, Rwanda using Sentinel-2 MSI data and ecosystem service bundles

• U-Net achieved good performance in urban imperviousness density extraction from Sentinel-2 MSI data. • The hybrid approach using Random Forest and U-Net improved urban land cover classification accuracies. • The expansion of built-up areas resulted in high demand for provisioning and regulating ecosystem services. • The decrease in cropland and forest negatively affected ecosystem service supply. • The framework using Sentinel-2 MSI data and ES bundles is effective for urbanization monitoring and environmental impact. Rapid urbanization in developing countries often results in uncontrolled urban growth. In order to support sustainable urban development, reliable and up-to-date information on urban land cover changes and their environmental impact is needed. In this study, we aim at evaluating the potential of Sentinel-2 (S-2) Multispectral Instrument (MSI) data for urban land cover change monitoring and for analyzing resulting impacts on Ecosystem Services (ES) provision in Kigali, Rwanda. Land cover classification into eight distinct urban classes (84% overall accuracies, 0.8 Kappa) was performed on data from 2016 and 2021 using a hybrid approach combining Random Forest with a U-Net-based impervious surface segmentation that improved the delineation of urban areas. The bi-temporal land cover maps were then analyzed regarding landscape structure using Landscape Metrics (LM). Ecosystem service bundles were derived for both years and their changes were summarized. Service providing areas were further evaluated in terms of changes in spatial attributes and structure of patches. ES were aggregated into eight bundles and grouped into provisioning, regulating and supporting services. The bundles were further analyzed using a matrix spatially linking landscape units with service supply and demand budgets. The results illustrated that three urban development scenarios can be distinguished including infill through housing and infrastructures development in core urban areas, urban sprawl in fringe zones and the development of urban patches at distant locations intercepted by cropland. The results revealed that the changes in LM negatively affected ES supply mainly through a decrease in cropland and forest. The expansion of built-up areas resulted in a high demand for provisioning and regulating services, especially food and water provision, surface runoff mitigation and erosion control. This is the first study demonstrating that detailed monitoring of urbanization and resulting environmental impacts can be performed with open access S-2 MSI data in Sub-Saharan Africa. Moreover, the framework developed in this study has the potential to be transferred to other Sub-Saharan cities.

Detection and characterization of coastal tidal wetland change in the northeastern US using Landsat time series

Coastal tidal wetlands are highly altered ecosystems exposed to substantial risk due to widespread and frequent land-use change coupled with sea-level rise, leading to disrupted hydrologic and ecologic functions and ultimately, significant reduction in climate resiliency. Knowing where and when the changes have occurred, and the nature of those changes, is important for coastal communities and natural resource management. Large-scale mapping of coastal tidal wetland changes is extremely difficult due to their inherent dynamic nature. To bridge this gap, we developed an automated algorithm for DEtection and Characterization of cOastal tiDal wEtlands change (DECODE) using dense Landsat time series. DECODE consists of three elements, including spectral break detection, land cover classification and change characterization. DECODE assembles all available Landsat observations and introduces a water level regressor for each pixel to flag the spectral breaks and estimate harmonic time-series models for the divided temporal segments. Each temporal segment is classified (e.g., vegetated wetlands, open water, and others – including unvegetated areas and uplands) based on the phenological characteristics and the synthetic surface reflectance values calculated from the harmonic model coefficients, as well as a generic rule-based classification system. This harmonic model-based approach has the advantage of not needing the acquisition of satellite images at optimal conditions (i.e., low tide status) to avoid underestimating coastal vegetation caused by the tidal fluctuation. At the same time, DECODE can also characterize different kinds of changes including land cover change and condition change (i.e., land cover modification without conversion). We used DECODE to track status of coastal tidal wetlands in the northeastern United States from 1986 to 2020. The overall accuracy of land cover classification and change detection is approximately 95.8% and 99.8%, respectively. The vegetated wetlands and open water were mapped with user's accuracy of 94.6% and 99.0%, and producer's accuracy of 98.1% and 93.5%, respectively. The cover change and condition change were mapped with user's accuracy of 68.0% and 80.0%, and producer's accuracy of 80.5% and 97.1%, respectively. Approximately 3283 km2 of the coastal landscape within our study area in the northeastern United States changed at least once (12% of the study area), and condition changes were the dominant change type (84.3%). Vegetated coastal tidal wetland decreased consistently (~2.6 km2 per year) in the past 35 years, largely due to conversion to open water in the context of sea-level rise.

A Novel Efficient Method for Land Cover Classification in Fragmented Agricultural Landscapes Using Sentinel Satellite Imagery

Updated and accurate land cover maps are essential and crucial for sustainable crop production and efficient land management. However, accurate and efficient land cover mapping is still a challenge for agricultural regions with complicated landscapes. This study proposed a novel spectral-phenological based land cover classification (SPLC) method to identify the land cover for fragmented agricultural landscapes, with less requirement of ground truth data. The SPLC method integrated a pixel-based support vector machine (SVM) algorithm for cropland and various non-cropland classification, and a phenology-based automatic decision tree algorithm for identification of various crop types. It was then tested and applied in two typical case areas (i.e., Jiyuan in the upstream and Yonglian in the downstream) of Hetao Irrigation District (Hetao) in the upper Yellow River basin (YRB), northwest China. The field survey sampling data and the regional visual interpretation maps were jointly used to evaluate the accuracy of land cover classification. Results indicated that stable phenological rules can be established for crop identification even with complex planting patterns, and the SPLC method performed well in land cover mapping in case areas. Four high-accuracy land cover maps were produced for Jiyuan in 2020 and 2021, Yonglian in 2021, and Hetao in 2021. The overall accuracies (OA) can reach 0.90–0.94 based on evaluation with abundant ground truth data, and land cover maps agreed well with the visual interpretation maps in space. Overall, the case application validated the applicability and efficiency of the SPLC method in land cover mapping for regions with fragmented agricultural landscapes, and also implied the potential use in other similar regions.

Application of normalized difference vegetation index in classifying land cover change over Bangli regency by using Landsat 8 imagery

The information on land cover changes is very important in regional spatial planning. Remote sensing technology can minimize the cost and time in analyzing land cover changes. Normalized Difference Vegetation Index (NDVI) is a vegetation index that combines red and near-infrared channels so that it can provide approximate information about land cover in an area. The objective of this study is to extract land cover change information from Landsat 8 images based on NDVI values in Bangli Regency in 2015 and 2021. The classification method used to estimate the type of land cover is supervised classification. The results reveal the decrease of the land cover in the category of water body, sand, dry land/soil, rice fields, and vegetation, which are 1.62%, 14.14%, 7.93%, 8.63%, and 2.45%, respectively, while an increase in the settlement category by 30.12%. The overall accuracy of land cover classification result based on NDVI value is 86.54%.

A hierarchical category structure based convolutional recurrent neural network (HCS-ConvRNN) for Land-Cover classification using dense MODIS Time-Series data

Hierarchical classification of land cover can be used to describe the Earth’s surface with different scales and properties. However, existing studies have rarely considered hierarchical information for land-cover classification, and have ignored dependencies in the hierarchical structure. In this study, we propose a hierarchical category structure-based convolutional recurrent neural network (HCS-ConvRNN). The HCS-ConvRNN method constrains the input through the leaf node of the hierarchical structure based input layer, and then constructs the dependencies among different layers in a top-down manner, in order to classify the pixels into the most relevant classes in a layer-by-layer manner. A total of 219 Moderate Resolution Imaging Spectroradiometer (MODIS) images of China from 2015 to 2017, at a 5-day interval, were used in the reported experiments. It is shown that: 1) the results of HCS-ConvRNN have rich spatial details; 2) the accuracy at each level of HCS-ConvRNN is better than that of MOD12Q1; and 3) generally HCS-ConvRNN can obtain a better classification performance than other networks such as the convolutional neural network (CNN) and gated recurrent unit (GRU). In summary, the proposed HCS-ConvRNN method can effectively achieve hierarchical land cover classification, and has the potential for accurate land cover classification at a large scale.

AGFP-Net: Attentive geometric feature pyramid network for land cover classification using airborne multispectral LiDAR data

Accurate land cover (LC) classification plays an important role in ecosystem protection, climate changes, and urban planning. The airborne multispectral LiDAR data are increasingly used for high-resolution and accurate LC classification tasks. However, most of the existing methods lack of the comprehensive extraction of the spatial geometric structure features, and ignore the fusion of multi-scale extracted features. In this paper, a point-wise deep learning-based method is proposed for LC classification based on airborne multispectral LiDAR data. We present a novel convolution operator to efficiently extract the spatial geometric structure features, called attentive graph geometric moments convolution (AGGM Convolution). Besides, to fuse the extracted multi-scale features, we propose a feature up-sampling module and construct a feature pyramid to integrate the features with different scales. The proposed method was evaluated using multispectral LiDAR data acquired with an airborne Teledyne Optech Titan system. In comparison with the previously developed state-of-the-art point cloud segmentation models, the proposed method behaves superiorly with an overall accuracy of 96.9% and a Kappa index of 0.950 on the test scenes. The quantitative assessments demonstrate that the proposed method performs effectively and efficiently in land cover classification tasks.