Tree species composition mapping with dimension reduction and post-classification using very high-resolution hyperspectral imaging

ABSTRACTA number of clear issues are pertinent when considering whether, or not, to use a remotely sensed dataset. We evaluate these issues here by comparing an aerial hyperspectral image at 1.5 m geometric resolution that comprises 128 narrow bands within a spectral range between 400 nm and 1,000 nm as well as a nine-band Landsat 8 image at 30.0 m geometric resolution. We therefore applied Random Forest (RF) and Support Vector Machine (SVM) classifiers utilizing different input data sets to determine the best thematic accuracy for both types of images by involving all possible bands and then minimized them using variable selection and dimension reduction via Minimum Noise Fraction (MNF). We then compared Landsat images to an aerial hyperspectral one. The results of this analysis revealed that band selections based on variable importance and MNF-transformation improved thematic accuracy assessed as Overall Accuracy (OA). Results reveal a 1.00% improvement in OA via variable selection as 59 bands instead of 128 bands and a 1.50% via MNF-transformation of the hyperspectral image. This improvement was 4.52% in the Landsat image when using a MNF-transformation compared to the best performances without transformation or variable selection. Data also showed that application of Landsat spectral range on hyperspectral bands resulted in different outcomes; specifically, SVM resulted in a 91.50% OA while RF resulted in 95.50% OA. Landscape ecology results show that use of the Landsat image provided fewer land cover patches and that differences encompassed 6.30% of the whole area. We therefore conclude that Landsat data can be used with a number of limitations for accurate ecological mapping.

Combining data of different types and from different sources for the classification of tree species has gained popularity recently, but training models on such datasets often requires more computational demands and does not always result in higher accuracy due to feature redundancy and irrelevance. Thus preprocessing data using dimensionality reduction (DR) methods can be employed to improve the classification accuracy and reduce computations. The objective of this research is to investigate and compare tree species classification performance for different classification algorithms [naive Bayes (NB), logistic regression (LR), random forest (RF), and support vector machine (SVM)], combined with various DR methods (correlation-based feature selection filter, information gain, wrapper methods, and principal component analysis). Two primary datasets are used—QuickBird and LiDAR, as well as derived topography data. When DR is used prior to classification, the NB classifier had a significant improvement in accuracy. SVM and RF had the best classification accuracy without DR. The overall accuracies (OA) of SVM and RF are 88.2% and 87.2% (kappa 0.84 and 0.83), respectively, followed closely by LR (OA: 84.8%, kappa: 0.79) and more distantly by NB (OA: 79%, kappa: 0.72). It is recommended to use SVM and RF without DR or NB with DR for tree species classification.

For dimensionality reduction (DR) of a hyperspectral data cube or band selection, it is desirable to have one method that is suitable for all remote sensing applications. However, in reality this is not possible. A specific remote sensing application requires a specific DR or band selection method that best suits it. In this paper, the evaluation and comparison of three DR methods‐namely, principal component analysis (PCA), wavelet, and minimum noise fraction (MNF)‐and one band selection method were conducted. Based on the experiments, the following was observed. For endmember extraction, the PCA DR, wavelet DR, and band selection found all five endmembers. However, the MNF DR missed one endmember. For mineral detection, the MNF DR produced a map that is closest to the true map when compared with the other DR methods and band selection method. For classification, the PCA DR produced the highest classification rates whereas the other methods yielded less classification rates.

Accurate and reliable information on tree species composition and distribution is crucial in operational and sustainable forest management. Developing a high-precision tree species map based on time series satellite data is an effective and cost-efficient approach. However, we do not quantitatively know how the time scale of data acquisitions contributes to complex tree species mapping. This study aimed to produce a detailed tree species map in a typical forest zone of the Changbai Mountains by incorporating Sentinel-2 images, topography data, and machine learning algorithms. We focused on exploring the effects of the three-year time series of Sentinel-2 within monthly, seasonal, and yearly time scales on the classification of ten dominant tree species. A random forest (RF) and support vector machine (SVM) were compared and employed to map continuous tree species. The results showed that classification with monthly datasets (overall accuracy (OA): 83.38–87.45%) outperformed that with seasonal and yearly datasets (OA:72.38–85.91%), and the RF (OA: 81.70–87.45%) was better than the SVM (OA: 72.38–83.38%) at processing the same datasets. Short-wave infrared, the normalized vegetation index, and elevation were the most important variables for tree species classification. The highest classification accuracy of 87.45% was achieved by combining RF, monthly datasets, and topography information. In terms of single species’ accuracy, the F1 scores of the ten tree species ranged from 62.99% (Manchurian ash) to 97.04% (Mongolian Oak), and eight of them obtained high F1 scores greater than 87%. This study confirmed that monthly Sentinel-2 datasets, topography data, and machine learning algorithms have great potential for accurate tree species mapping in mountainous regions.

Hyperspectral images (HSI) contain a wide range of information, the most prominent technology for observing the earth. However, using an original HSI high-dimensional datacube, the classification task faces significant challenges since it has a high computational cost. As a result, dimensionality reduction is indispensable. A dimension reduction method has been introduced in this paper, including feature extraction and feature selection to obtain feature subsets. Minimum Noise Fraction (MNF) is a popular feature extraction method for HSI, requiring a high computational capability. We propose a segmented MNF that divides the complete HSI into groups utilising normalised cross-cumulative residual entropy (nCCRE). An nCCRE-based feature selection is also employed to improve the quality of the chosen features using the max-relevancy min-redundancy measure. The support vector machine (SVM) classifier is used on two real HSI to evaluate the efficiency of the extracted subsets.

Accurate and reliable lithological mapping through satellite-borne remote sensing data and image classification approaches has a critical role since it can automatically and promptly identify lithological units over large areas. Most available Pixel-Object Based comparative classification studies have been applied to land use land cover (LULC) studies; however, this research aims to evaluate and compare the performance of these digital classification methods in the field of geological mapping in semi-arid areas, by integrating spectral bands and neo-bands, particularly the Minimum noise fraction (MNF) and the principal component analysis (PCA), of Sentinel-2A satellite imagery, to map the southern of Skhour Rehamna which is located at the western Moroccan Meseta. The analysis results from two different methods, namely, pixel-based image analysis (PBIA) with k-nearest neighbour (K-NN) and Random Forest (RF) machine learning algorithms (MLAs), and Geographic Object-Based Image Analysis (GEOBIA) were assessed and compared. PBIA method involved selection of training areas whether it was k-NN or RF MLAs, and produced lithological maps that exhibit “salt and pepper” effects as well as problems associated to delineating accurate lithological boundaries, while GEOBIA approach involved multi-resolution segmentation step where scale, shape and compactness parameters should be adjusted as accurate as possible, in order to segment the image into homogeneous and meaningful regions so that the resulted samples were classified using Standard Nearest Neighbour algorithm. Therefore, the resulting lithological maps were assessed by comparing both techniques using confusion matrix, overall accuracy (OA) and Kappa coefficient (K). The results show that the GEOBIA approach had higher overall agreement (83.46% OA and 0.76 K) than RF (81.92% OA and 0.72 K) and k-NN (80.79% OA and 0.70 K) PBIA approaches. Overall, the results clearly indicate the potential of GEOBIA technique for lithological mapping applications to produce more realistic maps.

Considering the form diversity of tree species composition in the Bagong Mountain National Forest Park of China, we mapped tree species utilizing Machine Learning Algorithms (support vector machines (SVM) and random forest (RF) classifiers) based on the OHS-2 hyperspectral satellite image by different datasets which combined spectral information and hyperspectral-derived vegetation indices (VIs) for improving tree species classification and explored the best performance of them. To verify the improvement, the results of physically-based spectral classifiers (spectral angle mapper (SAM) and maximum likelihood (ML) classifiers) were applied to compare with the results of machine learning algorithms. The results indicated an overall accuracy of 94.01%, 96.08%, 82.9% and 79.3% for SVM, RF, SAM and ML classifiers of the best performance using different datasets. Highest accuracies resulted from two machine learning algorithms classifiers; SVM and RF compared to SAM and ML classifiers. Although SVM outperformed RF when using all hyperspectral bands and VIs, the overall accuracy of the RF classifier is higher when compared to the SVM classifier using VIs combined selected features. Meanwhile, the RF classifier performed better than SVM after removing the redundancy of spectral data in training samples. Moreover, the machine learning algorithms successfully classified a small number of tree species (Cedrus deodara and Pterocarya stenoptera C. DC.) in the study area, but the physical spectroscopy-based method failed to classify these species. Such integration strategy improved the effectiveness of enhancing the accuracy of tree species classification and mapping their distribution on broad spatial and temporal scales using machine learning algorithms and hyperspectral imagery.

One of the purposes of hyperspectral remote sensing is to differentiate and identify the materials present on the Earth's surface by the spectral behavior of each object in the different regions of the electromagnetic spectrum. Such differentiation and identification can be accomplished through different image classification algorithms. However, there is no perfect classifier, since every algorithm has labeling errors. With the advent of orbital and aerial images of very high spatial and spectral resolution, the recognition of the materials present in urban environments is increasingly accurate. Thus, we thoroughly study different methodologies to identify the algorithm that presents the best results in the characterization of urban objects. The hyperspectral image used in the present study represents an area over Houston University - Texas and its surroundings, containing 48 spectral bands, with a spatial resolution of 1 meter and spectral range of 380 nm to 1050 nm. For the identification of 21 classes present in the study area, this paper analyzes two different classification methods: Deep Learning and Random Forest. To improve classification accuracy, performed the feature extraction. To obtain such preliminary results we used tools available in specific software as Normalized Difference Vegetation Index (NDVI), Minimum Noise Fraction (MNF), Principal Component Analysis (PCA) and Soil Adjusted Vegetation Index (SAVI). The image segmentation was performed using two different methods known as Multiresolution Segmentation and Spectral difference. Multiresolution segmentation needs parameters related to form and compactness. The best results were obtained with the values of form = 0.7 and compactness = 0.5, besides the scale of 10. From this, samples of all classes contained in the study area were selected for the training of the algorithms. This step is of paramount importance, as sample collection directly impacts the result of the classifications. After performing these steps, the information obtained from sample collection is entered into the data mining software (WEKA 3.8) to train the classification algorithms. The analysis of the results was performed by cross-validation, thus obtained the confusion matrix, calculated the Overall Accuracy (OA) and Kappa Index. The classification by the Random Forest method had an overall accuracy of 84.72% and a Kappa Index of 0.83. In turn, the Deep Learning algorithm had an overall accuracy of 81.32% and a Kappa index of 0.80. In this case, the classification by the Random Forest method presented better results for the hyperspectral image classification than the Deep Learning method. The accuracy difference obtained between the methods is not considered significant, so it is suggested for future work to analyze other complementary issues such as processing time.

This study utilizes multi-sensor satellite images and machine learning methodology to analyze urban impervious surfaces, with a particular focus on Nanchang, Jiangxi Province, China. The results indicate that combining multiple optical satellite images (Landsat-8, CBERS-04) with a Synthetic Aperture Radar (SAR) image (Sentinel-1) enhances detection accuracy. The overall accuracy (OA) and kappa coefficients increased from 84.3% to 88.3% and from 89.21% to 92.55%, respectively, compared to the exclusive use of the Landsat-8 image. Notably, the Random Forest algorithm, with its unique dual-random sampling technique for fusing multi-sensor satellite data, outperforms other machine learning methods like Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), Classification and Regression Trees (CARTs), Maximum Likelihood Classification (Max-Likelihood), and Minimum Distance Classification (Min-Distance) in impervious surface extraction efficiency. With additional satellite images from 2015, 2017, and 2020, the impervious surface changes are tracked in the Nanchang metropolitan region. From 2015 to 2021, they record a notable increase in impervious surfaces, signaling a quickened urban expansion. This study observes several impervious surface growth patterns, such as a tendency to concentrate near rivers, and larger areas in the east of Nanchang. While the expansion was mainly southward from 2015 to 2021, by 2021, the growth began spreading northward around the Gan River basin.

This study evaluates the utility of the ensemble framework of feature selection and machine learning (ML) models for regional landslide susceptibility mapping (LSM) in the arid climatic condition of southern Peru. A historical landslide inventory and 24 different landslide influencing factors (LIFs) were prepared using remotely sensed and auxiliary datasets. The LIFs were evaluated using multi-collinearity statistics and their relative importance was measured to select the most discriminative LIFs using the ensemble feature selection method, which was developed using Chi-square, gain ratio, and relief-F methods. We evaluated the performance of ten different ML algorithms (linear discriminant analysis, mixture discriminant analysis, bagged cart, boosted logistic regression, k-nearest neighbors, artificial neural network, support vector machine, random forest, rotation forest, and C5.0) using different accuracy statistics (sensitivity, specificity, area under curve (AUC), and overall accuracy (OA)). We used suitable combinations of individual ML models to develop different ensemble ML models and evaluated their performance in LSM. We assessed the impact of LIFs on ML performance. Among all individual ML models, the k-nearest neighbors (sensitivity = 0.72, specificity = 0.82, AUC = 0.86, OA = 78%) and artificial neural network (sensitivity = 0.71, specificity = 0.85, AUC = 0.87, OA = 79%) algorithms showed the best performance using the top five LIFs, while random forest, rotation forest, and C5.0 (sensitivity = 0.76–0.81, specificity = 0.87, AUC = 0.90–0.93, OA = 82–84%) outperformed other models when developed using all twenty-four LIFs. Among ensemble models, the ensemble of k-nearest neighbors and rotation forest, k-nearest neighbors and artificial neural network, and artificial neural network and rotation forest outperformed other models (sensitivity = 0.72–0.73, specificity = 0.83–0.84, AUC = 0.86, OA = 79%) using the top five LIFs. The landslide susceptibility maps derived using these models indicate that ~2–3% and ~10–12% of the total study area fall within the “very high” and “high” susceptibility. The obtained susceptibility maps can be efficiently used to prioritize landslide mitigation activities.

As a kind of cultural art, calligraphy and painting are not only an important part of traditional culture but also has important value of art collection and trade. The existence of forgeries has seriously affected the fair trade, protection, and inheritance of calligraphy and painting. There is an urgent need for the efficient, accurate, and intelligent technical identification method. By combining the advantages of material attribute recognition and imaging detection of hyperspectral imaging technology with the powerful feature expression ability and classification ability of the convolutional neural network, it can greatly improve the comprehension efficiency of calligraphy and painting identification; meanwhile, in order to reduce the redundancy and the amount of parameters in the method of directly using the hyperspectral image, an objective convex dimensionality reduction method should be used for compressing the original hyperspectral image before deep learning. Based on this, we propose a kind of deep learning method to classify author and authenticity based on the multichannel images obtained by minimum noise fraction (MNF) dimensionality reduction to calligraphy and painting hyperspectral data, and its core is the 2D-CNN or 3D-CNN model with the basic network of “4 convolution layers + 4 pooling layers + 2 full-link layers.” The experimental results show that the identification accuracy of the 2D-CNN calligraphy and painting identification with MNF pseudocolor image mosaic as input and the 2D-CNN calligraphy and painting identification with multichannel MNF dimensionality reduced images direct as input have high accuracy, while the 3D-CNN calligraphy and painting identification with multichannel MNF dimensionality reduced images direct as input not only maintains excellent identification accuracy but also has better learning convergence (step number) and stability compared with the 2D-CNN model. Especially, the 3D-CNN identification accuracy of calligraphy and painting's author and authenticity on the test set can reach 93.2% and 95.2%, respectively.

Maize lodging is a prevalent stress that can significantly diminish corn yield and quality. Unmanned aerial vehicles (UAVs) remote sensing is a practical means to quickly obtain lodging information at field scale, such as area, severity, and distribution. However, existing studies primarily use machine learning (ML) methods to qualitatively analyze maize lodging (lodging and non-lodging) or estimate the maize lodging percentage, while there is less research using deep learning (DL) to quantitatively estimate maize lodging parameters (type, severity, and direction). This study aims to introduce advanced DL algorithms into the maize lodging classification task using UAV-multispectral images and investigate the advantages of DL compared with traditional ML methods. This study collected a UAV-multispectral dataset containing non-lodging maize and lodging maize with different lodging types, severities, and directions. Additionally, 22 vegetation indices (VIs) were extracted from multispectral data, followed by spatial aggregation and image cropping. Five ML classifiers and three DL models were trained to classify the maize lodging parameters. Finally, we compared the performance of ML and DL models in evaluating maize lodging parameters. The results indicate that the Random Forest (RF) model outperforms the other four ML algorithms, achieving an overall accuracy (OA) of 89.29% and a Kappa coefficient of 0.8852. However, the maize lodging classification performance of DL models is significantly better than that of ML methods. Specifically, Swin-T performs better than ResNet-50 and ConvNeXt-T, with an OA reaching 96.02% and a Kappa coefficient of 0.9574. This can be attributed to the fact that Swin-T can more effectively extract detailed information that accurately characterizes maize lodging traits from UAV-multispectral data. This study demonstrates that combining DL with UAV-multispectral data enables a more comprehensive understanding of maize lodging type, severity, and direction, which is essential for post-disaster rescue operations and agricultural insurance claims.

This paper presents an approach for high spatial resolution hyperspectral image analysis in an applied task of river water condition assessment. The method allows the detection of algal blooms or water pollution by foreign substances. High-resolution hyperspectral images were obtained using a hyperspectrometer mounted on a small unmanned aerial vehicle. A difference between the spectra of river parts with varying intensity of algal blooms was demonstrated. Water samples were taken, and chemical analysis confirmed the varying levels of magnesium and calcium across all samples, corresponding to the intensity of algal blooms in the water. Several machine learning-based classification algorithms and vegetation indices were considered for classifying water areas with varying intensities of algal blooms. The effectiveness of machine learning algorithms compared to vegetation indices was shown. In addition, to improve the performance of the most effective classification algorithms, a comparison of several dimensionality reduction approaches based on spectral channel selection was carried out.

Sensors support machine learning

Lithological studies and geological unit mappings are generally applicable to many fields of natural resource management. Relatively suitable aquifers have been formed in complex formations in northwest Shahrood due to the presence of carbonate rocks as well as erosion and tectonic forces in the region. This study aims to identify and separate the calcareous formations that can form karst aquifers in the study area. As a result of erosion and tectonic forces, the rocks of the region exhibit spectral fluctuations, making it difficult for mapping geological formations using multispectral images. Therefore, Landsat 8 satellite-based images were processed by adopting the minimum noise fraction (MNF), independent component analysis (ICA), and band ratio (BR). The indices of calcareous formations and shale formations were created by the BR through the spectral behavior of pure pixels. Moreover, the support vector machine (SVM) and maximum likelihood (ML) were employed for classification. The SVM classifier proved more capable of classification than the ML classifier, and the transform ICA outperformed the MNF in the separation of formations. The lithological maps were extracted using the SVM with an overall accuracy (OA) of 68.04%. Furthermore, a method decision tree (DT) was employed to improve the classification accuracy. The DT classifier was then utilized to reclassify lithological maps that were classified by SVM and ML through morphological characteristics and indices of formations The DT classifier improved the lithological map accuracy by 10%. The boundaries of calcareous formations were extracted from non-calcareous formations with an accuracy of 93%, and the regional constructions were separated with an accuracy of ∼80 % . Finally, the lithological map was developed with a kappa of 0.734 and an OA of 78.59%.