Plant species recognition has long been an important topic in forest remote sensing and ecology. In recent years, Convolutional Neural Networks (CNNs) have become the main approach for image-based plant recognition. However, most existing approaches use the collected data to build a CNN classifier with single-time training without taking into account the condition of updating of data. If the CNN model continues to learn new data, it may encounter a catastrophic forgetting issue, where the model fails to recognize the classes in the old data. To address this, a learning strategy known as Continual Learning (CL), or life-long learning, has been developed. CL techniques can maintain a certain level of discriminative ability from dynamically-updated data. In this paper, we adopt a hyperspectral plant canopy image dataset of 120 species as the materials to conduct an experiment based on CL classification. We divide the dataset into 4 tasks to simulate the scenario of obtaining new class images in batches. In each task, we apply the CL techniques using techniques known as knowledge distillation and memory replay to prevent catastrophic forgetting issue. In the experimental setting, we also consider two different band selection methods for data dimensionality reduction. The results show that using CL techniques significantly improve higher classification accuracy as well as maintain the classification ability for old classes.

This study was conducted to reduce the bias associated with using ICESat-2 data in canopy height estimation and mapping by integrating the features of Sentinel-1&2. The parameters derived from Sentinel-1 and Sentinel-2 images were used for the canopy height modeling through the gradient boosting regression model. The ICESat-2-based canopy height served as the training data while the ALS data was used for validating the accuracy of the predicted canopy height product. The result showed a low correlation between the ICESat-2 and ALS-based canopy height. Nevertheless, the accuracy of the predicted canopy height was increased, from 14.08m to 13.89m based on RMSE and from 55.62% to 54.86% based on PRMSE, when the features of Sentinel-1 was integrated with the features of Sentinel-2. Further, low canopy height values tend to be overpredicted while high canopy height tend to be underpredicted. Thus, canopy height correction models can be developed to augment issues of low accuracy associated with canopy height modeling using ICESat-2.

Plant species recognition is an important topic in forest remote sensing. In the past, people have used laboratory plant images to design different recognition methods. Some applied machine learning algorithms to perform species recognition on the plant RGB images, while others started to use deep learning (DL) approaches, such as convolutional neural network (CNN), to improve recognition performance. However, two issues arise: Firstly, many species with similar appearance are difficult to be correctly classified by CNN under the limited spectral information of RGB imaging. Secondly, the existing CNN-based classification models are designed for particular datasets and thus are not suitable for plant images. To tackle these issues, this paper proposes a revolutionary framework that combines hyperspectral imaging (HSI) and DL technologies to perform plant species classification with a large number of species. We collected canopy images of 100 plant species via a Visible-NIR hyperspectral camera, and built an plant canopy dataset which consists of 3250 training images and 3250 test ones. Furthermore, we designed a lightweight CNN model that utilizes both 3D CNN and 2D CNN modules to implement spectral information fusion and spectral-spatial features extraction, called Hybrid-CNN. Experimental results show that Hybrid-CNN achieved at least 98% in overall accuracy rate and 0.98 in Kappa coefficient, and significantly outperformed the reference classification models.

Forest resources inventory and monitoring are essential for efficient management of natural resources. Remote sensing-based mapping technology provides geospatially explicit information on forest ecosystems. Generally, the attributes of forests derived from moderate satellite images show spatial and classification inaccuracy, introducing uncertainty in formulating forest dynamics and misleading inappropriate management plans. This study examines the efficiency of low-cost UAVs in generating orthoimages for differentiating bamboo from broadleaf species in forest succession. The results show that the generated very high-resolution orthoimage can reveal detailed spatial features of both bamboo and tree crowns, providing an excellent opportunity to differentiate individual bamboo crowns from trees and therefore help derive bamboo expansion and degradation over the forest.

k nearest neighbor (kNN) algorithm, a non-parametric and one of the machine learning techniques had been popularly used for stand-level AGB estimation. kNN algorithm has two distinct parameters which are concerning the distance and the number of nearest neighbors (k) and the distance of target plots with reference plots. This study was conducted to assess the effects of the number of k, reference plots density, and distance of target plots with reference plots on the performance of the kNN algorithm in estimating stand-level AGB information. Using the surface reflectance of the Sentinel-2 image, the effects of the number of k (3, 5, 8, 11), reference plots density (high=3932 pixels, low=656 pixels), and the distance of the target and reference plots (2km, 5km, and 10km) to the prediction error of the kNN algorithm were assessed. The performance of the kNN was assessed using the RMSE and RMSPE as evaluation criteria. The 3-way factorial ANOVA was used to determine the main and interaction effects of the variables. The results showed that the number of k, reference plot density, and the distance of target plots with reference plots have a significant effect on the prediction error of the kNN algorithm. The PRMSE increases with the increasing distance and decreased with the increased number of k. Further, the PRMSE was low when the reference plot density was high. The PRMSE can be improved by as much as 13% when the reference plot density is sufficient and when an optimal number of k and distance are used for the parameters.

Deep learning techniques have been demonstrated with a pronounced performance in diverse object recognition and classification fields. An accurate distribution map of trees provides sufficient information on forest ecosystems and underpins the need for the sustainable management of forests. For classification with a smaller number of reference datasets, applications of the CNN technique with transfer learning and fine-tuning process were reported to be suitable in the literature. This study applied the Mask R-CNN technique to tree species mapping for a hemiboreal mixed forest in Hokkaido, Japan. An orthoimage with 25 cm resolution generated via airborne RGB image was used for this study. The Mask R-CNN model was derived from a 5-ha reference image and evaluated by a 1-ha test image. The experimental results show that a moderate accuracy (F1 score = 0.72) can be achieved for the forest with six dominant species in this study. The accuracy measure changed dramatically in the species, ranging from 0.20 to 0.94. This accuracy appeared to be a function of the sample size in the reference dataset for machine learning via high-resolution airborne RGB images. In this work, the species with sample images of more than 200 seem sufficient to achieve an F1 score greater than 0.90.

Global warming is becoming more and more obvious hastening the effect of climate change and at the same time affecting the growth of plants. Under global warming, the growth of vegetation is affected by climatic factors such as temperature and rainfall, which have also led to changes in the structure of the timberline and its distribution along the altitudes. In order to observe the trend of timberline in Taiwan under global warming, this study focused on Yushan peaks as study site, which represents the Taiwan's alpine ecosystem, using multi-temporal Landsat 8 satellite images that were acquired in 2013, 2016 and 2019 and meteorological data. The satellite images were atmospherically corrected using the FLAASH and pansharpened using the NNDiffuse technique. The different land cover in Yushan namely forest, low vegetation, bareland, and river were classified using the SVM supervised classification. The changes in the distribution of trees in different periods with the changes in the meteorological data were identified, classified, analysed and compared. The results of the study show that the current impact of global warming to climate change and on Taiwan's timberline trend is caused mainly by temperature. The rising temperature overtime had caused the timberline to shift towards the higher altitudes.

Agricultural crops monitoring covering large-scale areas with high-efficiency is of great importance in sustainable agriculture management to combat the demand for agricultural crops brought by the exponential growth of population worldwide. The study was conducted to assess the promising potential of segmenting cloud free and high-resolution images for watermelon fruit monitoring using the combination of Unmanned Aerial Vehicle (UAV) orthophotographs and Support Vector Machine (SVM) supervised machine learning model. The digital images of the watermelon field were captured using UAV DJI Mavic 2 Pro through Pix4D operation and were orthorectified to generate an orthophotograph of the study site using the Agisoft Metashape software. After the segmentation using the edge segmentation and full lambda algorithm, the segmented objects were classified using the SVM technique with 150 and 100 segmented blocks as training samples representing the fruit of the watermelon and the non-fruit, respectively. The result revealed that use of high-resolution digital imagery and the application of machine learning techniques could be a tool to a wide area accounting of agricultural crops as indicated by the overall classification accuracy of 91% and a Kappa coefficient of 0.79.

Fuel loading is a critical parameter for wildfire management. A spatially explicit map of fuel loadings can help highlight wildfire risk and guide people to ignite fires more carefully and help to minimize the possibility of heat conduction over forest lands. In contrast to the airborne lidar based methods, a geospatial integrated data based algorithm was proposed to derive surface fuel loading map of a mountainous forest. Results showed that our method is able to produce a map of fuel loadings using biological parameter (normalized value of forest-type code) and topographic parameters (classified slope code, classified aspect code, and reciprocal of degree slope) at a reasonable level of accuracy. The prediction bias for the test dataset was at a level of RMSE=3.42 ton/ha and RMSE%=41 %. The estimation bias at each of the plot was averaged 2.3524±2.3528 ton/ha and the minimum and maximum bias was -13.25 and 6.47 ton/ha respectively. A natural logarithm transformation can help to improve the significance of each coefficient at the probability level of 0.05 while keeping the performance of the SFL model.

A low cost but accurate forest monitoring tool with is necessary for sustainable forest management. Lidar data is quite expensive and is challenging when it comes to multitemporal data collection. Unmanned-Aerial-Vehicle (UAV) remote sensing is low cost and has flexibility in high frequency data collection with a satisfactory image resolution net. Focused on forest growth evaluation, this study investigates the potential of integrating UAV and Airborne Laser Scanner (ALS) technology in the evaluation of tree volume in high stem-density forest stands. ALS data is used to provide base state of digital surface model (DSM) and digital elevation model (DEM). The UAV aerial photographs collected in a series of inventory were used to generate multitemporal DSM of forest stands by Structure from Motion (SfM) technique. With the ALS-DEM and UAV-SfM-DSM data, this study has shown a satisfied result in retrieving tree volume of a Hinoki (Chamaecyparis obtusa) forest. The estimation accuracy was RMSE=0.695m and RMSEP=3.96%. The performance indicates that the approach of UAV-SfM-DSM with ALS-DEM has high potential within forest stock evaluation.

A series of hyperspectral and chlorophyll experiments was conducted to gather WorldView multispectral data and derive an estimation model of chlorophyll concentration (CHLS) of tree leaves. The multispectral effective chlorophyll indicators (MECIb and MECIc) derived from the reflectance of Red, Red edge and Blue or Coastal bands of the WorldView data appeared to be able to achieve an estimation with around 38% of PRMSE, which is almost 4 times greater than the value of the measure (1-R2) of the CHLS-MECI models. Such uncertainty in tree leaves chlorophyll estimation was duplicated in both multiple regression models determined by ordinary least squares method and ridge method using the spectral signal of Yellow, Red, and Red edge bands. In contrast, a ridge regression model using the reflectance of both NIR bands, the normalized difference vegetation index of Yellow and Red bands (NDVIYR) as well as Red edge and Yellow bands (NDVIREY) was able to reduce the uncertainty of CHLS estimation significantly. In summary, the ridge model was able to increase estimation accuracy to a level of RMSE=0.35 mg/g and PRMSE=22%. The major strength of the ridge model was observed in a reduction of the estimation error of fresh leaves by 65% of the PRMSE of the CHLS-MECI models. However, there was no improvement observed to the CHLS estimation of water-stressed leaves.

This paper explores the feasibility of deriving multispectral-based effective chlorophyll indicators (MECIs) for foliage chlorophyll concentration (CHLS) estimation. An average fusion method was applied to simulate the multispectral reflectance of the WorldView-3 sensor using hyperspectral data. With the experimental data of CHLS and predictors derived from multispectral reflectance, a series of linear regression analyses were carried out to derive appropriate models for CHLS estimation. Accuracy measures of RMSE and PRMSE were used to evaluate the model performance. Results showed that the coastal-band based MECI (MECIc) and the blue-band based MECI (MECIb) were able to achieve an RMSE of 0.5657 mg/g and 0.5943 mg/g as well as a PRMSE of 36% and 38% respectively. Using the Red edge and Yellow reflectance based NDVI (NDVIREY) as a predictor, the model can reduce uncertainty and achieve an estimation of 0.4089 mg/g and 26% for RMSE and PRMSE respectively. The prediction error made by the CHLS-NDVIREYmodel and the CHLS-MECI model were 11% and 60% larger than 0.38 mg/g the RMSE of hyperspectral-based CHLS-ECI model. In summary, NDVIREYwas able to achieve a better prediction at around a level of 75% accuracy (1-PRMSE) and therefore is able to be an effective indicator of CHLS for forest monitoring.

This paper proposed an efficient height slicing method for detecting trees using a canopy height model (CHM). A digital forest simulator was proposed to randomly generate trees by a 2-dimensional Gaussian probability density function. Details such as the location, crown radius, and height of a tree are automatically generated via the parameters of location, spread blob, and amplitude of a two-dimensional Gaussian function. An index of hit rate was used to evaluate the detection power of the algorithm and the indices RMSE and PRMSE were used to evaluate the estimation accuracy of tree height and crown radius. The proposed algorithm was able to detect trees at 100% and 80% hit rate at a stand density of less than 600 trees per hectare and this then gradually decreased to 85% and 70% as stand density increased to 1000 trees per hectare for artificial forest and natural forest respectively.

This study explored the feasibility of mapping the canopy closures of mangrove forest in the Mekong Delta of Vietnam using Sentinel-2 multispectral composite image. Forest canopy closures were determined in accordance with the level of volume stocks. A method of object-based support vector machine classifier was first applied to derive LULC and then to differentiate the canopy closure over the mangrove forest. Results showed that object-based SVM classification was able to achieve an accuracy of kappa of around 0.73 which is around 0.2 higher than the kappa of a pixel-based SVM classification. However, there was a level of around 11%-24% commission rate and omission rate in the rich, medium, and poor classes of canopy closure. Further research to improve the performance of canopy closure classification is needed in order to obtain more accurate information for management planning.

A mitigation named reducing emission from deforestation and forest degradation (REDD+) has been developed as one of the most cost-effective strategies to tackle climate change due to forest loss. Timely and accurate forest data are required for REDD+ programs to succeed. A series of multitemporal Landsat imageries obtained during the period from 1995 to 2015 were used to evaluate deforestation in Central Kalimantan REDD+ boundaries. By using support vector machine (SVM) classification method, the LULC types was mapped with a kappa coefficient greater than 0.8 facilitated a reliable interpretation of LULC changes. In 1995, the forest occupied the total area at around 72.6% while it decreased dramatically to a level of 59.5% and 53.8% in 2005 and 2015. The deforestation in the REDD+ project area in the two decades from 1995–2005 and 2005–2015 was 321.2 and 131.1 thousand hectares.

Height structure of a forest stand is a generalized probability-density-function (PDF) model of the stand height which is the average height of dominant and co-dominant trees over a forest stand. An appropriate model of height structure is able to support precise estimation of forest carbon stocks as it is integrated with volumetric parameters at the level of sample plots and forest stands. This paper developed a protocol of deriving forest canopy height distribution models using ICESat GLAS footprint data. Results showed that an ordinary stand-height (OSH) model was able to predict stand height with an RMSE and RMSE% around 3.5 m and 25% at footprint level using the metrics waveform extent, leading edge extent, and trailing edge extent. Based on the footprint-level stand height estimates, a height structure model was derived as a log-logistic PDF (LL-PDF) distribution which is similar to the one derived from airborne lidar data. However, the shape and scale parameters of the GLAS-derived model were significantly greater than those of the lidar-derived model. The distribution of the GLAS-derived height structure showed a higher kurtosis and smaller data range. It is inferred that the OSH model determined stand height at footprint level has a discrepancy of estimation. This is expected from a model that is capable of retrieving crucial signals of footprint-level stand height.

This paper applied a Logistic-Gaussian complex signal model (LGM) to restore spectroradiometer-obtained canopy SWIR reflectance signals of red cypress, an old-growth conifer species. This species distributes within a temperate forest in Taiwan over a wide range of altitude from 1000 to 2800 m. A large amount of reflectance data of trees in a variety of light environments was used for this study. Although the reflectance curve of tree crowns is decreased as the directly incident light is blocked by shadow of uphill tree crowns, the particular pattern of tree reflectance curve in such a noisy region remained similar. Results showed that the pronounced noise in the region from 1350 to 1410 nm can be removed using the LGM complex signal model. In other words, the reflectance signals of trees in such spectral areas can be restored successfully. Briefly, the research revealed that the model shows acceptable ability to fix the noise problem in the water-sorption spectral wavelengths under a variety of incident radiance. This is valuable for the use of remotely sensed data for continuous monitoring of water stress in forest canopy.

Pollock model was one of the popular template matching algorithms, which constructs a three-dimensional description of each individual tree-crown envelope. Unfortunately, it is always challenging to determine parameters since each tree has different grown condition. This paper proposed an iterative least square fitting algorithm to form a Pollock model best fitting each individual tree for tree detection and crown delineation. The proposed algorithm provides a solution to avoid difficulties of finding the best set of parameters for each tree crown. As a result, it improves the detection rate and decreases the computing time as shown in the experimental studies. The proposed least square fitting algorithm is more suitable for practical applications.

Multi-level morphological active contour algorithm (MMAC), was proposed in [3-4], to incorporate mathematical morphology for tree detection and crown delineation with the average prediction error around 4% in mountainous areas. However, two major drawbacks come with the design of MMAC algorithm, which are huge computational complexity and certain level of false alarm or omission rate. The paper provides a solution for two major problems by proposing a gradient vector flow (GVF) snake based MMAC algorithm. The GVF snake model took advantage of the gradient vector force of the LiDAR image and expand the boundary of each blob to approximate the real crown of each forest candidate. According to the experimental studies, the proposed algorithm could help improve the computational complexity and detection rate of the original MMAC algorithm effectively.

During the last decade unmanned aerial systems (UAS) have been intensively applied to create 3-dimensional models of land surface features for various applications. Although, UAS data can be jointly used with airborne LiDAR data to generate a canopy height model (CHM) of forest stands, it is rare to find research concerned the generation of forest CHM using only UAS data. This paper investigate a suitable method to classify UAS point cloud to create CHM data for forest inventory. Results showed that ground points over building areas, open land, and forestland can be successfully collected by appropriate terrain angles which define a threshold value of the angle between a point, its projection on the plane of a triangle, and the closest vertex of a TIN surface model. A conservative threshold value of 5 degrees was suggested due to its allowing critical ground points whilst excluding crown points being collected. The UAS-derived CHM was evaluated with an RMSE accuracy of 0.01, 0.20, and 0.42 m for road, buildings, and trees respectively.

Recently, some global-scale DEM products, such as GTOPO30, ETOPO1, SRTM and ASTER GDEM have been published for geoscience applications. The latest product, ALOS DEM was announced to be available for a global coverage in 2016. This study examined the performance of ALOS-DEM in describing accurate morphometric and volumetric measurement of land features. A comparison was made on basis of DEM and DSM data of airborne full-waveform LiDAR data. Results showed that ALOS DEM is more approximately in reality an ALOS DSM which reveals the ground envelop surface rather than the ground bare surface. The differences between ALOS DEM and LiDAR DSM are mainly from 0 to 2.75 m with a standard deviation of 1.58 m. The differences between ALOS DEM and LiDAR DEM give a bias of as large as 20m, mostly located at the areas with abrupt change of relief and mainly in the north-facing slopes. This is probably due to ALOS sensor's geometry in corresponding to its looking-direction. The stream networks derived from both ALOS DEM and LiDAR DEM are in good agreement. It is suggested that further studies on methods for assessing geomorphometric changes in landform structures should be developed and compared.

This paper develops band weighting spectral methods, which are wSAM and wSID, for detection of pesticide residues on vegetables. Since the water content of vegetables has significant impact on the measured spectrum, the proposed band weighting measures are able to suppress the effect of water content to enhance detectability of of pesticide residue detection. Compared to the traditional band selection techniques, there are three advantages. First of all, it does not require determining the number of bands to be selected. Second, the proposed methods assigned a weight to each band based on the amount of pesticide information. Third, the band weighting method could help reduce the effect of undesired signal, which is the water content in our case. The experimental study further demonstrates the utilities of our proposed band weighting methods.

Multi-level Morphological Active Contour algorithm (MMAC) had been proposed to effectively increase recognition rate of individual tree in mountainous areas. However, it was specifically designed for LiDAR CHM data only, which make the algorithm incompatible with any other type of remote sensing data. To relieve constraints of MMAC this manuscript proposed a spectral-based MMAC (SB-MMAC), which retains the framework of MMAC by replacing height information from LiDAR CHM model with spectral information from multispectral images. The proposed SB-MMAC is comprised of two stages, seed blobs detection and modified active contour model. The experimental study further demonstrated the utility of SB-MMAC.

Forests in Taiwan contain a diverse variety. Remote sensing data could offer information of large area sampling, but it is very challenging to automatically detect tree and delineate three crown in remote sensing data. Recently a algorithm, called multi-level morphological active contour algorithm (MMAC), has been proposed to address these issues in [1]. It combines a multi-level morphological approach with the active contour model. However, this algorithm comes with a price, which is huge computational complexity, to prevent it from being implemented practically in medium- or large-scale images. This manuscript aimed to accelerate MMAC by developing a parallel processing approach using reconfigurable computing platform with field-programmable gate arrays (FPGA). The experimental evaluation indicated that the proposed architecture could provide around 40% acceleration with simple implementation in hardware.

An individual tree within a forest stand will have its height and diameter growth restricted by the influence of neighboring trees. This is because trees in close proximity compete for resources and space to enable growth. In this paper, the position of trees, tree height (LH), tree crown radius (LCR), and growth competition index (LCI) were extracted from a light-detection-and-ranging (LiDAR)-based rasterized canopy height model using the multilevel morphological active-contour algorithm. The diameter and volume of individual trees are tested and validated to be an exponential function of those LiDAR-derived tree parameters. The best LiDAR-based diameter estimation model and volume estimation model were tested as significant with anR2value of 0.84 and 0.9 and evaluated with an estimation bias of 8.7 cm and 0.91 m3, respectively. Results also showed that LH and LCR are positively related to the LiDAR-derived diameter at breast height (DBH) and the LiDAR-derived volume of individual trees in a forest stand, whereas LCI is negatively related. The proposed algorithm of individual tree volume estimation was further applied to predict the volume of three sample plots in mountainous forest stands. It was found that the LVM could be used to predict an acceptable volume estimate of old-aged forest stands. The estimation bias, i.e., percentage RMSE (RMSE%), is averaged at around 4% using the LiDAR metrics lnLH, LCI, and LCR, whereas the RMSE% increases to 50% if only lnLH is applied. Results suggest that LCI is an important regulation factor in the estimation of forest volume stocks using LiDAR remote sensing.

Support vector machine (SVM) has received considerable interest in hyperspectral image classification. In order to make SVM work effectively one challenge is selection of training samples. In supervised classification it is generally done by random sampling for cross validation where two issues must be addressed. One is how many training samples required to allow SVM to produce good performance and the other is how to deal with random selections of training samples which produce inconsistent results. This paper presents a new type of SVM, called iterative SVM (ISVM) to address these two issues. The idea is to implement an SVM iteratively in such a way that the sample size is not necessarily to be large while the random sampling issue can be also resolved. To substantiate the utility of ISVM Purdue data is further used for experiments.