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

Abstract

Tree species’ composition of forests is essential in forest management and nature conservation. We aimed to identify the tree species structure of a floodplain forest area using a hyperspectral image. We proposed an efficient novel strategy including the testing of three dimension reduction (DR) methods: Principal Component Analysis, Minimum Noise Fraction (MNF) and Indipendent Component Analysis with five machine learning (ML) algorithms (Maximum Likelihood Classifier, Support Vector Classification, Support Vector Machine, Random Forest and Artificial Neural Network) to find the most accurate outcome; altogether 300 models were calculated. Post-classification was applied by combining the multiresolution segmentation and filtering. MNF was the most efficient DR technique, and at least 7 components were needed to gain an overall accuracy (OA) of > 75%. Forty-five models had > 80% OAs; MNF was 43, and the Maximum Likelihood was 19 times among these models. Best classification belonged to MNF with 10 components and Maximum Likelihood classifier with the OA of 83.3%. Post-classification increased the OA to 86.1%. We quantified the differences among the possible DR and ML methods, and found that even > 10% worse model can be found using popular standard procedures related to the best results. Our workflow calls the attention of careful model selection to gain accurate maps.

Deep learning‐based training data augmentation combined with post‐classification improves the classification accuracy for dominant and scattered invasive forest tree species

“…The output is also similar to the PCA, the largest explained variances belong to the first 5–10 principal components (PCs) and then the contribution of the PCs' is decreasing. MNF was applied successfully as input data in several previous studies (Burai et al., 2019; Hamada et al., 2007; Likó et al., 2022; Zhang & Xie, 2012).…”

“…MRS was applied on the classified image having the highest OA (Fig. 2), similarly to our previous study (Likó et al., 2022), but we applied a more sophisticated version. The scale parameter was tested with 10 and 100 different levels, in ascending order from 1 to 10 with 1 spacing (L10–L100 levels), and from 0.1 to 10 with 0.1 spacing (L01–L100 levels).…”

“…It has three main parameters that can be set: scale, shape and compactness (Drǎgut ß et al, 2010). We selected the optimal shape and compactness parameters based on the preliminary tests (with the aim to find the tree crowns); thus, 0.5 compactness and 0.3 shape values were used (Lik o et al, 2022).…”

“…Spatial resolution (ground sampling distance, GSD) can also be an important factor when choosing the most suitable platform: UASs' hyperspectral sensors usually have a GSD of 0.1–0.5 m depending on the flight height, while in case of aircraft GSD can be 1–2 m. Currently, only a few hyperspectral satellite data sources are publicly available: the PRISMA (Vangi et al., 2021), the EnMap (Guanter et al., 2015; Stuffler et al., 2007) and the DESIS (Farmonov et al., 2023; Müller et al., 2016); all three systems featuring 30 m GSD. For tree species mapping, the appropriate pixel size is 1 m (e.g., Brovkina et al., 2015; Burai et al., 2019; Likó et al., 2022; Merentitis et al., 2014), thus favoring UASs‐ and aircraft‐based data gathering in such applications.…”

“…Different methods have been developed in order to improve accuracy, such as using textural information, spectral indices, ancillary data, visual interpretation, smoothing or post‐classification (e.g., Bhosale et al., 2014; El‐Hattab, 2016; Likó et al., 2022; Manandhar et al., 2009; Thakkar et al., 2017), but there is no standard methodology to choose one. The selection or development of the most effective procedure is highly dependent on multiple factors i.e., the target object (plant), the initial dataset and the obtained results before performing the post‐classification procedure.…”