84 research outputs found
Deriving a distribution model of forest canopy height at stand level using icesat glas full-waveform data
No full textA Novel Effective Chlorophyll Indicator for Forest Monitoring Using Worldview-3 Multispectral Reflectance
No full textA Generalized Logistic-Gaussian-Complex Signal Model for the Restoration of Canopy SWIR Hyperspectral Reflectance
No full textThe continuum of the SWIR (short-wave infrared) signals from 1320 to 1650 nm contains valuable information for effectively diagnosing water, chlorophyll, and nitrogen content. The SWIR spectra of in situ spectroradiometric data and airborne spectrometric images are frequently contaminated by significant noise. Based on a Logistic-Gaussian complex signal model (LGCM), the noise-free signals at 1330–1349 and 1411–1430 nm wavelengths can provide critical bases for restoring the 1350–1410 nm wavelength signals for a single point of data. This paper proposes a generalized LGCM (GLGCM) technique to expand the ability of LGCM to process large data with variant reflectance values. A 12-year-old red cypress plantation located in a central Taiwan temperate forest was selected for this study. Hundreds of reflectance spectra of tree crowns were obtained using an ASD FR Spectroradiometer. The in-laboratory blank test showed that the GLGCM technique was able to achieve sufficient performance with an RMSE (root mean square error) of 0.0015 ± 0.0005 and 0.0011 ± 0.0005 for the front-edge and end-edge signal bases respectively, and 0.0014 ± 0.0006 in between the two signal bases. A significant level of noise between −0.2 and 0.4 was successfully removed from the in situ contaminated reflectance in the 1350–1410 nm wavelengths. The estimation bias for the signals of front-edge and end-edge bases was low, averaging 0.0031 ± 0.0003 and 0.0032 ± 0.0012. The consistency between the blank test and the in situ experimental results indicates that the GLGCM technique has potential in using batch processing to fix the problem of the noisy SWIR spectra in spectroradiometeric data and also airborne spectrometric images
Applying a logistic-Gaussian complex signal model to restore surface hyperspectral reflectance of an old-growth tree species in cool temperate forest
No full textPlant Species Recognition Based on Plant Canopy Hyperspectral Images and Convolutional Neural Networks
No full textPlant species recognition has been an important research issue. In the past, people identified species via the appearance features of plants, such as roots, stems, leaves, flowers, fruits, etc. However, this method requires a wealth of knowledge and experience. In recent years, many people started to use the Machine Learning (ML) techniques such as support vector machines, k-nearest neighbor algorithms, etc. with RGB plant images to recognize plants. However, most of these methods only adopt leaf image for recognition, and the recognition performance relies on the quality of feature extraction. With the rise of deep learning (DL), Convolutional Neural Network (CNN) began to be the representative method in image processing and computer vision issues. CNN can realize automatic feature learning and classification through the convolutional layers, which is much more efficient than traditional ML methods. However, the spectral information contained in RGB images is limited. People began to use other imaging techniques to enhance the recognition ability.
In this thesis, we use Hyperspectral Imaging (HSI) to produce plant images, and propose a CNN called Hybrid Convolutional Neural Network (Hybrid CNN) to realize accurate plant image recognition. This approach utilizes the spectral information of hyperspectral and the ability to integrate spatial information of CNN to identify plants. Our Hybrid-CNN is composed of 3DCNN and 2DCNN network modules, in which 3DCNN is used to enhance the feature fusion in both spectral and spatial dimension. We collect plant canopy images of 100 different plant species and produce a hyperspectral image dataset for this topic. In terms of experimental settings, we use several different band selection (BS) methods to reduce the dimensionality of hyperspectral images, and observe the impact of using different BS methods on the final results. The experimental results show that the accuracy rate of more than 99% can be achieved by Hybrid-CNN in all BS settings. Also, the Hybrid-CNN outperforms the existing LtCNN used for the compared method. This explains that the combination of hyperspectral imaging and deep learning has great potential in plant species recognition
Application of Deep Learning-based Spectral Reconstruction Technology in Tissue Identification and Insect-Eroded Area Detection on Lotus Leaf Images
No full textUsing image processing technology to detect plant diseases is one of the important research issues. Since many plant diseases in early stage are difficult to detect in visible spectrum, we must observe the phenomenon of early diseases through the spectral images which can provide the information in near infrared spectrum, and then implement early prevention and control to reduce the future economic losses caused by pest diseases. In recent years, multispectral (MSI) sensors have been widely used in aerial imagery which can provide 2 to 5 near infrared light bands, making it easier to find insect-eroded areas. Although it has brought many benefits over RGB images, the spectral information is still insufficient for accurate detection. To overcome this issue, we can use hyperspectral (HSI) camera to capture more accurate spectral information. However, the high price of HSI devices is still a considerable expense for researchers. Under such circumstances, using deep learning (DL) technology to reconstruct RGB or MSI images into HSI images became a new research issue.
In this paper, the spectral reconstruction technique is utilized to perform tissue identification and insect-eroded areas detection in lotus leaf images. Firstly, we use 150-band 470-900nm hyperspectral image data of lotus leaves as the materials to train a CNN model that can reconstruct the RGB images to 150-band HSI. In addition to analyze the reconstructed quality, we also perform image analysis via CEM, NCLS, LSOSP, 1D-CNN, and SVM algorithms. Later, we apply such technique to 10-band aerial MSI images of lotus fields in which a CNN model that can covert 10-band MSI to 150-band HSI is trained. The results show that using the reconstructed HSI image helps to improve the performance of image analysis in both quantitative and qualitative analysis. For example, using the reconstructed lotus field images, the overall classification accuracy rate produced by each classifier can reach more than 95%. In the results of detection algorithms, the area under the ROC curve can mostly reach more than 0.8. Besides, for the detection of erosion invisible areas of using reconstructed indoor lotus leaf images, the erosion regions can be more clearly found than using original HSI image
Adjacency effect to hyperspectral signatures in field spectroscopy: The influence of pixel size and polarization of light energy
No full textAnalyzing Land use Land Cover and Deforestation in the REDD+AREA, Central Kalimantan Province, Indonesia
No full textIdentifying forest ecosystem regions for agricultural use and conservation
No full textABSTRACT Balancing agricultural needs with the need to protect biodiverse environments presents a challenge to forestry management. An imbalance in resource production and ecosystem regulation often leads to degradation or deforestation such as when excessive cultivation damages forest biodiversity. Lack of information on geospatial biodiversity may hamper forest ecosystems. In particular, this may be an issue in areas where there is a strong need to reassign land to food production. It is essential to identify and protect those parts of the forest that are key to its preservation. This paper presents a strategy for choosing suitable areas for agricultural management based on a geospatial variation of Shannon's vegetation diversity index (SHDI). This index offers a method for selecting areas with low levels of biodiversity and carbon stock accumulation ability, thereby reducing the negative environmental impact of converting forest land to agricultural use. The natural forest ecosystem of the controversial 1997 Ex-Mega Rice Project (EMRP) in Indonesia is used as an example. Results showed that the geospatial pattern of biodiversity can be accurately derived using kriging analysis and then effectively applied to the delineation of agricultural production areas using an ecological threshold of SHDI. A prediction model that integrates a number of species and families and average annual rainfall was developed by principal component regression (PCR) to obtain a geospatial distribution map of biodiversity. Species richness was found to be an appropriate indicator of SHDI and able to assist in the identification of areas for agricultural use and natural forest management
- …
