Comparison of crop classification in single-period data of Xiguang No. 1 05 (Tianxianpei No.) and Sentinel-2 data.

Background Introduction

Identifying and differentiating the distribution of various crop types using remote sensing imagery enables refined management and macro-level monitoring of agricultural production. This monitoring provides data support for pest and disease early warning. Furthermore, remote sensing classification results, combined with meteorological, hydrological, and soil information, provide a scientific basis for agricultural yield estimation, disaster loss assessment, and food security decision-making. Its significance lies in promoting the transformation of traditional agriculture towards digitalization, informatization, and intelligentization, improving agricultural production efficiency and resource utilization, reducing manual survey costs, and achieving dynamic monitoring and precise management of large-scale agricultural areas. In addition, crop classification results can provide crucial foundational data for national agricultural policy formulation, arable land protection, and carbon sequestration assessment, which has profound implications for ensuring food security and promoting sustainable agricultural development.

Multispectral imagery typically contains a small number of broad bands (e.g., 4–13), primarily reflecting the overall characteristics of ground features in key bands such as visible and near-infrared light. It is suitable for macro-scale ground feature classification and monitoring, such as land use and vegetation cover. Hyperspectral imagery, on the other hand, is used in…Continuous sampling in tens to hundreds of narrow bandsHyperspectral data can capture the fine reflectance characteristics of ground features in the spectrum, enabling the identification of subtle differences that cannot be distinguished by multispectral methods. For example, variations in different crop varieties, vegetation physiological states, or mineral composition can all be accurately captured in hyperspectral data. By comparing the two, the impact of spectral resolution on the accuracy of ground feature identification and feature extraction capabilities can be revealed, clarifying the optimal data selection scheme for different tasks.

Principles and methods

This experiment was conducted using data from the Xiguang-1 05 satellite (also known as Tianxianpei) and Sentinel-2. First, the Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), and Enhanced Vegetation Index (EVI) were calculated for both hyperspectral and multispectral data. Then, for the hyperspectral data, a modified firefly algorithm was used to select bands with high correlation for band dimensionality reduction. Finally, Support Vector Machine (SVM) was used to classify crops in both the hyperspectral and multispectral data.

Results Analysis

Figure 1. Hyperspectral classification results

Figure 2 Multispectral classification results

In this crop classification experiment, three indices—NDVI, RVI, and EVI—were used, along with bands highly correlated with crops. The SVM algorithm was employed to minimize the influence of the classifier. Figures 1 and 2 show the classification results. The single-period hyperspectral classification results were superior to the multispectral classification results. Multispectral images suffer from blurred boundaries, numerous spots, and lower accuracy, while hyperspectral images provide finer-grained and more stable ground feature identification capabilities.

In land cover classification, hyperspectral data has significant advantages over multispectral data. Hyperspectral imagery records the reflectance characteristics of land cover across multiple continuous narrow bands, enabling a more detailed characterization of the spectral differences between different crop types on the land surface. It often exhibits distinguishable inter-spectral features in hyperspectral space, thus significantly improving classification accuracy and class separability. Using hyperspectral data, spectral feature matching methods can achieve fine differentiation and high-confidence identification of complex land cover types. In contrast, traditional multispectral data has fewer bands and wider bandwidth, providing only limited spectral information. It is prone to spectral overlap between different land cover types, leading to blurred classification boundaries and misclassification. Furthermore, multispectral data is unstable in complex surface environments (such as mixed land cover types, shadows, and soil background interference), making it difficult to effectively capture subtle differences. Therefore, from the perspectives of classification accuracy, robustness, and interpretability, hyperspectral data has significant advantages in land cover classification tasks and is an important data source for achieving high-precision land surface information extraction.