Abstract: In this study, we automate tree species classification and mapping using field-based training data, high spatial resolution airborne hyperspectral imagery, and a convolutional neural network classifier (CNN). We tested our methods by identifying seven dominant trees species as well as dead standing trees in a mixed-conifer forest in the Southern Sierra Nevada Mountains, CA (USA) using training, validation, and testing datasets composed of spatially-explicit transects and plots sampled across a single strip of imaging spectroscopy. We also used a three-band ‘Red-Green-Blue’ pseudo true-color subset of the hyperspectral imagery strip to test the classification accuracy of a CNN model without the additional non-visible spectral data provided in the hyperspectral imagery.
Our classifier is pixel-based rather than object based, although we use three-dimensional structural information from airborne Light Detection and Ranging (LiDAR) to identify trees (points > 5 m above the ground) and the classifier was applied to image pixels that were thus identified as tree crowns. By training a CNN classifier using field data and hyperspectral imagery, we were able to accurately identify tree species and predict their distribution, as well as the distribution of tree mortality, across the landscape. Using a window size of 15 pixels and eight hidden convolutional layers, a CNN model classified the correct species of 713 individual trees from hyperspectral imagery with an average F-score of 0.87 and F-scores ranging from 0.67–0.95 depending on species.
The CNN classification model performance increased from a combined F-score of 0.64 for the Red-Green-Blue model to a combined F-score of 0.87 for the hyperspectral model. The hyperspectral CNN model captures the species composition changes across ~700 meters (1935 to 2630 m) of elevation from a lower-elevation mixed oak conifer forest to a higher-elevation fir-dominated coniferous forest. High resolution tree species maps can support forest ecosystem monitoring and management, and identifying dead trees aids landscape assessment of forest mortality resulting from drought, insects and pathogens. We publicly provide our code to apply deep learning classifiers to tree species identification from geospatial imagery and field training data
Wissenschaftlicher Artikel2022
A convolution neural network for forest leaf chlorophyll and carotenoid estimation using hyperspectral reflectance
Shuo Shi, Lu Xu, Wei Gong, Bowen Chen, Biwu Chen, Fangfang Qu, Xingtao Tang, Jia Sun, Jian Yang
Forest leaf chlorophyll (Cab) and carotenoid (Cxc) are key functional indicators for the state of the forest ecosystem. Current machine learning models based on hyperspectral reflectance are widely applied to estimate leaf Cab and Cxc contents at leaf scale. However, these models have certain accuracy for non-independent datasets but have poor generalization for independent datasets when they are used to estimate leaf Cab and Cxc contents. This fact limits that hyperspectral remote sensing completely replaces destructive measurements for leaf Cab and Cxc contents. Thus, the development of an estimation model with high accuracy and satisfactory generalization is necessary. Convolutional neural networks (CNNs) have certain accuracy and generalization in many domains, and have the potential to solve above-mentioned problem. Therefore, this study developed a CNN using one-dimensional hyperspectral reflectance, which aimed to improve the model’s accuracy and generalization in leaf Cab and Cxc content estimation at leaf scale. The proposed CNN was developed by three steps. First, in consideration of the correlation between leaf Cab and Cxc contents in natural leaves, 2500 physical data with leaf reflectance and corresponding Cab and Cxc contents were generated by leaf radiative transfer model and multivariable gaussian distribution function. Then, the proposed CNN was built by five strategies based on the architecture of the AlexNet. Finally, five-fold cross validation was performed with 70% of the physical data to determine the best strategy to develop the proposed CNN. These were executed to ensure the proposed CNN with the maximum accuracy and generalization. In addition, the accuracy and generalization of the proposed CNN were tested using a non-independent dataset and an independent dataset, respectively. The proposed CNN was also compared with back propagation neural network (BPNN), support vector regression (SVR) and gaussian process regression (GPR). Results showed that the best CNN could be developed with one input, five convolutional, three max-pooling and three fully-connected layers. Comprehensively considering the model’s accuracy and generalization, the proposed CNN was the best model for leaf Cab and Cxc content estimation compared with BPNN, SVR and GPR. This study provides a development strategy of CNN estimation model using one-dimensional hyperspectral reflectance at leaf scale. The proposed CNN could further promote the practical application of hyperspectral remote sensing in leaf Cab and Cxc content estimation.
