The correct classification of individual tree species (ITS) is of great importance for forest resource management. However, current ITS samples collection is time-consuming and laborious in field surveys, and the samples cannot be updated in time. In this study, a sample update method is proposed, which updates samples by detecting the spectral angular similarity of samples at different times at the same location, enabling the reuse of historical samples and timely update of samples. Experimental results showed that the classification accuracy could reach more than 90% by using DenseNet model and the updated samples, which proved that the sample update method could achieve sample update and sample reuse effectively. In addition, there was significant improvement in the classification accuracy of the sample set using multispectral and texture information fusion, proving the effectiveness of information fusion in increasing classification information.

Accurate individual tree crown (ITC) detection is crucial for forest management, ecological studies, and sustainable harvesting practices. This study comprehensively investigates ITC detection in a mixed-wood forest t characterized by overlapping crowns of varying shapes and sizes. We employ a deep learning approach combined with traditional treetop detection algorithms. Our methodology involves generating a confidence map using the U-Net model to estimate the probability of treetop occurrence at each pixel in a Canopy Height Model, followed by treetop detection. The results demonstrate that our proposed method, integrating U-Net with a blob detection algorithm, achieved a detection accuracy of 89% with an omission error of 0.13 and a commission error of 0.16, enhancing the results compared to traditional methods included in the study. Our approach provides an effective strategy for ITC detection in complex forest environments.

A decision-level fusion method was developed to classify eleven cover types in a sub-arctic ecosystem, including intertidal marsh, tundra heath, peat plateau, open fen, shrub-rich fen, wet fen, conifer swamp, shrub swamp, conifer, and lichen woodland using Sentinel-1 and Sentinel-2 data. An index was also designed to measure the classification uncertainty for individual pixels. Three classifiers based on Random Forest (RF) classification were first designed and carried out, and the classification results were then combined within the framework of Dempster-Shaffer (DS) theory. The developed method increased the overall accuracy by 2.5% based on the test samples and reduced the percentage of pixels with high uncertainty by 6.4%, compared with the feature-level fusion method.

Given the recent thrust to convert forests in the Clay Belt to agricultural land, there is a vital need to assess what the attendant effects on carbon and greenhouse gas emissions will be under these scenarios. Different carbon modellers are used to estimate the soil carbon stock under possible land management schemes following the conversion from forests. The use of higher resolution remotely sensed data as input to the carbon modellers is explored in this study. Comparisons are made of the emissions estimated from the models when more accurate data is input to the models vs. having approximate data as input. Finally, the total ecosystem carbon is evaluated to determine how the amounts sequestered as well as the greenhouse gas emissions which could be produced from the possible land conversions.

Land use dynamics as observed from geospatial data has had a variety of important applications in society. In the agriculture and forestry industry for instance, changes detected from satellite imagery are used to monitor crop growth, detect soil moisture, monitor deforestation, and forest fires. This paper highlights the methods developed to detect and map land cover and land use in the Clay Belt of Northern Ontario, Canada. Advances in data sensing technologies and spatial analysis have afforded the opportunity to conduct feature-based classification and vegetative classification of the landcover. Utilising remotely sensed data from multiple sources such as airborne LiDAR and satellite images of Landsat and Sentinel, more accurate land cover changes can be observed. These land cover maps of the Clay Belt area have applications in improving modelling of soil carbon and greenhouse gas emissions.

An updated treatment of deriving the gap probability for a discontinuous canopy, specifically for right cones, is presented with respect to the framework of Li & Strahler [1]. Rather than employing numerical integration, equations of path length for a right cone have now been formulated and utilized. Three cases of ray transmission through the right cone are elucidated, corresponding to a ray passing through the side, as well as exiting through the bottom for whether half-apex angle is less than or exceeds the zenith angle of illumination. Path length histograms have been calculated for various zenith angles, for an example conical tree. Between-crown gap probabilities have been computed for various zenith angles, and conform to results obtained by Li & Strahler [1] for a right conical canopy.

Spatial information of tree species composition of forest and urban vegetation is very important for forest protection and urban management. Tree species classification using remote sensing data is mainly conducted using such classification methods as SVM (Support Vector Machine) and Random Forest. Images used include multispectral/hyperspectral images, LiDAR (Light Detection And Ranging) data, or the combination of them. As a fast-growing and powerful tool having obtained state-of-the-art results in many remote sensing applications, Deep learning (DL) has a great potential in outperforming these existing classification methods and obtaining more accurate classification maps. In this work, three CNN models were employed for tree species classification at individual tree level, using the combination of high-resolution multispectral image (i.e., WorldView-2) and LiDAR data. Compared the traditional object-based classification using random forest (RF) and support vector machine (SVM), the 18-layer ResNet and the 40-layer DenseNet provided significant higher accuracies. The experimental results indicate the advantages of the two CNN models used for tree species classification.

Models for the prediction of soil texture for the Abitibi River Forest (ARF) region in the District of Cochrane in Ontario, Canada were created from environmental covariates generated from remotely-sensed data as soil formation factors. A novel approach of incorporating LiDAR (Light Detection and Ranging) retrievals for the entire study area to derive covariates of canopy height model (CHM) and gap fraction was investigated. CHM and gap fraction had high variable importance for the soil texture models fitted for the region, with CHM being the most important variable out of a set of 104 predictors, and gap fraction among the top predictors. Random forest (RF) and support vector machine with radial basis functions (SVM Radial) approaches were utilized for the soil texture classification. The inclusion of CHM and gap fraction with other environmental predictors improved upon the accuracy of soil texture models, with accuracy scores exceeding 0.7 and Cohen's kappa greater than 0.5. Prediction maps for soil texture were generated for the ARF study region.

The objective of this study was to improve individual tree crown delineation by fully exploiting the crown information exhibited in multi-wavelength LiDAR data. The data used in this study were obtained by an Optech's Titan instrument with three wavelengths: 532 nm, 1064 nm, and 1550 nm. Methods were developed to employ both spectral and structural information of tree crowns to separate crowns from other cover types and from each other. The methods were tested using a data set obtained over a study area in Toronto, Ontario, Canada. Preliminary results show that with both the canopy height model and intensities corresponding to the three wavelength, trees were distinguished from other cover types with a high accuracy and trees were separated from each other with a reasonable accuracy (based on visual observation).

An automatic algorithm for forest road identification and extraction was developed. The algorithm utilized Laplacian of Gaussian (LoG) filter and slope calculation on high resolution multispectral imagery and LiDAR data respectively to extract both primary road and secondary road segments in the forest area. Also, a hierarchical post process was designed to iteratively connect the road segments to form the final road map. The process was tested on Hearst forest, located in central Ontario, Canada. Based on visual examination against manual digitized roads, all the roads from the test area have been identified and extracted from the process..

This paper discussed an approach developed for the characterization of tree structures for the identification of Ash trees. The objective of this study was to support the effort in saving Ash trees from the attack by the Emerald Ash borer by identifying Ash trees from mobile LiDAR data. Using this data set, we employed a hieratical elimination approach to test if a tree was an Ash tree through a series of stages based on knowledge on the characteristics of an Ash tree. Preliminary results were promising, showing that our proposed methods had the ability to filter out non-Ash trees. We look to test more trees in the future and expand our filtering stages to assist in more accurate Ash tree identification.

This paper discusses an approach developed for tree species identification and subsequent health determination. Using LiDAR data's ability to provide complementary data to traditional optical imagery analysis methods, this paper introduces methods used to identify Ash trees within an urban landscape and classify their health. Using the ability of LiDAR to provide 3D information, tree types will first be distinguished, afterwards tree branching behavior will be used to classify the tree species. After identifying the tree species, the health will be determined.

Soil productive is critical to Earth surface vegetation management and monitoring with significant environmental and economic value. The human impact has played a very important role in agriculture and forest soil disturbances. In this study, very high spatial resolution multi-spectral imagery was investigated for its value and potential in soil treatment study. Individual tree height and DBH were retrieved from the imagery and was used to estimate the biomass, which is expected to be important indicator of ground soil treatment. Controlled ground experimental sites were used to validate the result and preliminary results have shown the biomass can't directly reflect the tree health condition. The normalized year difference would be a much better indicator for this study.

The Emerald Ash Borer (Agrilus planipennis, EAB) is one of the most destructive insects damaging all Ash species of the genus Fraxinus in Ontario, Canada. It is crucial to detect the EAB invasion as early as possible to allow possible treatments and reduce economic loss. The challenge is that there are limited symptoms indicating EAB infestation due to this pest's cryptic life stage. The current detection methods are all in-situ approaches, which are labour intensive and economically inefficient. In this study, the object oriented vegetation indices and texture information derived from Hyperspectral imagery were investigated to test the hypothesis that stressed Ash trees are more vulnerable to EAB and can be used to predict infestation levels.

The paper investigated the advantage of high density airborne LiDAR data for improving species classification of individual tree. The investigation is comprised of two stages, feature extraction and classification. Several feature metrics were derived from LiDAR data, most of which were to characterize the vertical structural properties of difference species. Some other metrics were calculated statistically from intensity and return number information. A supervised decision tree algorithm was applied on the extracted features to perform both feature selection and classification. Two classification themes were carried out: classification of coniferous and deciduous trees, and classification of five species. Experiment was conducted in Canadian boreal forests dominated by mature trees. The results demonstrated LiDAR derived vertical profile metrics are capable for species classification either to separate coniferous and deciduous or to separate multiple species. The best overall classification accuracy is 81.7% validated by using the test data from the same ecosystem as the training data.

A good similarity measure is very important to ensure accurate classification of vegetation species using hyperspectral remote sensing data. In this study, the effectiveness of the existing similarity measures, spectral angle mapper (SAM), Euclidean distance (ED), and spectral information divergence (SID) was evaluated using data measured by a field spectrometer. To overcome the limitations of the existing measures, a new metric was developed based on the concept of conditional entropy.

Surface albedo plays an important role in climate model simulations. Current climate models usually use simplified approaches to calculate albedo and can not take sub-grid heterogeneity into account. For heterogeneous land areas, retrieving large scale surface albedo by using current albedo characterization schemes can cause considerably spatial scaling bias. The scaling biases in the albedo estimation processes mainly result from overlooking sub-pixel variability of land surface characteristics and non-linear relationships between albedo and related parameters. The objective is to establish a new methodology to further reduce spatial scaling bias of surface albedo at coarse resolution. Contexture-based and texture-based methods were applied to remove spatial scaling bias. In addition, a new method, dealing with spatial variation of between and within land cover type, was proposed and applied. The results indicate that lumped albedo value can be considerably biased from the distributed albedo (about 20% on average). New proposed corrective algorithm is generally effective for heterogeneous boreal forest pixels.

Forest species classification is important for forest management and environment monitoring and protection. As the conventional methods are mainly based on the spectral signatures of forest canopies and the results are at stand level. With the high spatial resolution data, classification at individual tree level becomes achievable. The objective of this study is to develop a novel algorithm of tree crown segmentation for automatic classification at individual tree level. The approach uses the integrating laser scanning data with high spatial multispectral optical imagery. The automatic classification is composed of two stages: (1) tree crown segmentation, (2) species classification. The approach is applied to the test area consisting of both conifers and deciduous trees. The segmentation result shows good accuracy of crown delineation, and over 90% of the trees are segmented correctly.

In this study, vegetation cover type classification was investigated using CHRIS data over agricultural scenes acquired across the 2004 growing season. Spectral indices sensitive to crop chlorophyll content and leaf area index were first calculated from CHRIS nadir data in May, June and July. The seasonality of these indices was analyzed and employed to identify crop types in the study area. To further improve the classification accuracy, the angular signatures of the vegetation canopies were derived from the nadir and off- nadir data in the red and near-infrared band using the kernel-driven Ross_Thick and Li-Sparse model. The coefficients of the kernel-based model were then used for crop type classification, together with the spectral indices derived from nadir data. Preliminary results show that the additional angular information can slightly improve the classification accuracy.

Information on the distribution of forest species is critical to sustainable management of the forest resources. However, forest species structure information accuracy remains low. Most of the current multi-angle data algorithms are based on the satellite-borne or simulated datasets. In this study, we exploited the use of structural information derived from airborne data to compare the signatures from the ground survey. This investigation was carried out by using FIFEDOM (frequent image frames enhanced digital ortho-rectified mapping) data. This study used FIFEDOM data collected over the Algoma forest, Ontario, Canada, which has four dominant species, jack pine, black spruce, poplar and white birch. The accuracy of the radiometric and geometric multi-angle signature results were assessed against CASI (compact airborne spectrographic imager) data, which were collected at the same time and also compared to SPRINT model simulation results.

The Frequent Image Frames Enhanced Digital Orthorectified Mapping (FIFEDOM) camera was designed to provide a cost-effective remote-sensing method for accurate acquisition of forest information, such as spatial distributions of individual tree species and tree structures for forest monitoring and management. Compared with existing regular digital cameras, the FIFEDOM camera has several unique features as follows: (1) it can collect data not only in the visible bands (550 and 670 nm) but also in the near-infrared band (800 nm); (2) it has a frame rate of up to 3 frames/s with a frame size of 3500 times 2300; and (3) it has a wide angular field view with 150deg along track and 78.8deg across track. Its high frame rate and wide angular field view allow it to obtain a sequence of images that oversample ground target areas. The multiangle database and bidirectional reflectance signatures of forest canopies can be generated from the oversampled image data, which can be used to identify forest species and estimate tree structures. In addition, the multiframe highly overlapped FIFEDOM data can also be used to generate a very dense, high-quality, and reliable digital surface model. Effective methods for radiometric and geometric calibration of the FIFEDOM camera were developed in this paper. A data-acquisition campaign was carried out in 2004 over the Algoma boreal forest, Ontario, Canada. The FIFEDOM data were validated using the data acquired by the Compact Airborne Spectrographic Imager instrument, which was flown together with the FIFEDOM camera.

The FIFEDOM (Frequent Image Frames Enhanced Digital Ortho- rectified Mapping) camera was designed to provide a cost-effective remote sensing method for the accurate acquisition bidirectional surface signatures, which for forest scenes is expected to yield information related to spatial distributions of individual tree species and tree structure with application in forest monitoring and management. Compared with existing regular digital cameras, the FIFEDOM camera has several unique features: (1) it can collect image data not only in the visible bands (550 nm and 670 nm), but also in the near-infrared band (800 nm); (2) it has a frame rate of up to 3 frames per second with a frame size of 3500 x 2300; and (3) it has a wide angular field view with 150 degrees along track and 78.8 degrees across track. Its high frame rate and wide angular field view allow it to obtain a sequence of images that over-sample ground target areas. The multi-angle database and bi-directional reflectance signatures of forest canopies can be generated from the over-sampled image data, which can be used to estimate metrics of tree structural properties.

In this study, we used laboratory data to investigate the effects of multiple scattering between tree crowns and snow background on the linearity of mixtures of the reflectance of winter forest scenes. Several scenes were designed in the laboratory to simulate the natural forest winter landscape. Hyperspectral images of the designed scenes were acquired by the Compact Airborne Spectrographic Imager (CASI). Each scene was decomposed by linear spectral unmixing of the scene reflectance spectrum using sunlit crown, shaded crown, sunlit background, and shaded background as end members. The SPRINT canopy model was employed to evaluate the results of the linear spectra unmixing approach. Our results show that if a linear unmixing approach is used for the designed scenes, the errors in the fractions of end members are as high as 25 percent relative to the fractions obtained by the SPRINT model. This investigation suggests that non-linear mixture models may be needed to account for the multiple scattering between tree crowns and snow background.

In this study, impacts of vector quantization compression on prediction of leaf chlorophyll content of crops for the application to precision agriculture were evaluated. The compression algorithm tested in this paper is called successive approximation multi-stage vector quantization (SAMVQ). The hyperspectral data used were acquired by CASI over corn fields at L' Acadie experimental farm (Agriculture and Agri-Food Canada) during the summer of 2000. Nine zones in the corn fields with different fertilization levels (no fertilization, intermediate fertilization, and over-fertilization) were used to evaluate the difference between the leaf chlorophyll contents obtained from original and reconstructed reflectance data cubes. The root mean square errors (RMSEs) and the correlations between the chlorophyll content derived from the original data cube and that derived from the reconstructed data cubes were calculated in the nine zones. The spatial variability of chlorophyll content in the nine zones was also examined for the images of chlorophyll content created from the original and reconstructed reflectance data cubes. The results show that the chlorophyll content image created from the reconstructed reflectance data cube corresponding to SAMVQ with a compression ratio of 20 maintains good agreement with that derived from the original reflectance data cube in the nine zones in terms of estimated pigment mean and spatial variability. As a result, SAMVQ with a compression ratio of 20 is considered acceptable for the retrieval of crop chlorophyll content from CASI hyperspectral data for agriculture corn crops.

The Canadian Space Agency is investigating an onboard compressor for a hyperspectral satellite using its two innovative data compression techniques. It is essential to verify the quality of the compressed data and users' acceptability in terms of their remote sensing applications. Hyperspectral data cubes acquired by hyperspectral sensors such as casi, AVIRIS, Probe-1 and Hyperion were tested. Statistical hypothesis tests were used to assess if the means and variances in each spectral band of specified zones calculated from the reconstructed data cubes are significantly different from those calculated from the original data cube. Remote sensing end products, such as red edge, chlorophyll content and spectral unmixing were used to evaluate the compressed data. Preliminary test results show that hyperspectral data compressed using the two compression techniques at compression ratio up to 30:1 are acceptable in terms of statistical tests and remote sensing end products.

This paper systematically studies the interrelationship between surface bidirectional reflectance distribution function (BRDF) retrieval and atmospheric correction. The study uses the atmospheric correction scheme of the Moderate Resolution Imaging Spectroradiometer (MODIS) and angular sampling expected for MODIS and the Multiangle Imaging SpectroRadiometer (MISR) for different land cover types and optical depths of aerosols. The results show the following two points. 1) Even for a nonturbid atmosphere, the assumption of a Lambertian surface in atmospheric correction causes relative errors in the retrieved surface reflectances that average from 2 to 7% in the red and near-infrared bands, with worst cases showing errors of up to about 15% for turbid conditions. Consequently, it is necessary for improved accuracy to consider surface anisotropy in atmospheric correction. 2) Surface BRDF retrieval and atmospheric correction can be coupled in a converging iteration loop that improves the quality of atmospheric correction and subsequent BRDF retrievals. For example, performing two steps of the iteration loop is already sufficient to obtain mean relative errors of less than 1% in the retrieved surface reflectances even for an atmospheric aerosol optical depth of 0.4. As BRDF retrieval accuracies improve, so do bihemispherical albedo retrieval accuracies, with mean relative errors being 1-5% when using a Lambertian assumption and less than 1% after two iteration steps.