Hyperspectral Feature Selection for SOM Prediction Using Deep Reinforcement Learning and Multiple Subset Evaluation Strategies

Abstract

It has been widely certified that hyperspectral images can be effectively used to monitor soil organic matter (SOM). Though numerous bands reveal more details in spectral features, information redundancy and noise interference also come accordingly. Due to the fact that, nowadays, prevailing dimensionality reduction methods targeted to hyperspectral images fail to make effective band selections, it is hard to capture the spectral features of ground objects quickly and accurately. In this paper, to solve the inefficiency and instability of hyperspectral feature selection, we proposed a feature selection framework named reinforcement learning for feature selection in hyperspectral regression (RLFSR). Specifically, the Markov Decision Process (MDP) was used to simulate the hyperspectral band selection process, and reinforcement learning agents were introduced to improve model performance. Then two spectral feature evaluation methods were introduced to find internal relationships between the hyperspectral features and thus comprehensively evaluate all hyperspectral bands aimed at the soil. The feature selection methods—RLFSR-Net and RLFSR-Cv—were based on pre-trained deep networks and cross-validation, respectively, and achieved excellent results on airborne hyperspectral images from Yitong Manchu Autonomous County in China. The feature subsets achieved the highest accuracy for most inversion models, with inversion R2 values of 0.7506 and 0.7518, respectively. The two proposed methods showed slight differences in spectral feature extraction preferences and hyperspectral feature selection flexibilities in deep reinforcement learning. The experiments showed that the proposed RLFSR framework could better capture the spectral characteristics of SOM than the existing methods.

Estimation and Mapping of Soil Organic Matter Content Using a Stacking Ensemble Learning Model Based on Hyperspectral Images

“…On the one hand, the estimation model based on image pixel spectra and soil sample content is a data-driven framework with high requirements on the quality and quantity of the input data. Conversely, the input pixel reflectance values of space-borne hyperspectral images are easily affected by external factors in the acquisition, such as electromagnetic interference, cross-mixing issues among band channels, natural illumination, and topographic conditions [17][18][19], leading to biased pixel reflectance values and weakened response relationships between spectral reflectance and the SOM content. Additionally, the large number of spectral channels in hyperspectral images induces high collinearity and information redundancy among spectral data, reducing the quality of input data to a certain extent.…”

Enhancing Hyper-Spectral Image Classification with Reinforcement Learning and Advanced Multi-Objective Binary Grey Wolf Optimization

“…Reinforcement learning (RL), such as deep Q-learning (DQL), offers a different approach to HS image classification [27][28][29][30][31]. The classification of HS images through DQL represents a significant leap in the analytical capabilities of remote sensing (RS).…”

Quantitative analysis of potentially toxic elements in soil by XRF based on efficient reinforcement learning and sparse partial least squares

“…Presently, reinforcement learning variable selection is effectively used for soil quality assessment based on hyperspectral. 29 However, the structure of XRF spectra has a continuous character. The selection of valid variables by a spectral variable selection algorithm is usually based on the assumption that the property of interest is related to only a few spectral variables.…”