Ship identification based on FCLS pixel demixing

Background Introduction

In the fields of marine and inland waterway monitoring, vessel identification has always been a crucial application of remote sensing technology. Vessel distribution is directly related to shipping traffic management, maritime law enforcement, and national defense security, as well as fisheries resource monitoring, illegal fishing control, environmental protection, and emergency rescue. Traditional vessel monitoring methods primarily rely on radar and optical sensors, but these are susceptible to interference and have blind spots under complex sea conditions, cloud cover, or nighttime conditions. The emergence of hyperspectral remote sensing has brought new possibilities for vessel identification. In particular, the spectral unmixing method based on Fully Constrained Least Squares (FCLS) can separate water and vessel endmember components at the sub-pixel scale, providing a more robust and precise means for vessel identification in complex backgrounds. Therefore, hyperspectral vessel identification technology combined with FCLS is gradually becoming an important research direction in marine monitoring and intelligent remote sensing analysis.

The aim of this study is to fully leverage the spectral advantages of hyperspectral remote sensing and explore a ship identification strategy based on the FCLS (Fully Constrained Least Squares) unmixing method, thereby improving detection accuracy and stability in complex backgrounds. First, by treating the ship hull as an "abnormal endmember" in the aquatic environment, FCLS is used to decompose the pixel spectrum under abundance constraints. This allows for the quantification of the ratio of ship material to water components, overcoming the limitations of traditional image resolution for small-scale ship identification. Combining spatial morphological constraints and residual anomaly analysis, this study aims to construct a multi-feature fusion ship detection criterion to achieve robust identification of ships at different scales and under different background conditions. Ultimately, the research goal is not only to improve the automation and accuracy of ship identification but also to provide reliable data support and technical solutions for maritime traffic supervision, fisheries law enforcement, marine resource management, and emergency rescue.

Data and Methods

1. Data

This case study uses hyperspectral imagery of the Mediterranean strait (Bonifacio Strait) between Corsica, France and Sardinia, Italy, taken by Xiguang-1 05 satellite (Tianxianpei) on July 15, 2025, with atmospherically corrected L2C class reflectance.Rate data is used to identify ships in the area using the FCLS algorithm.

Table 1. Data Introduction

2. FCLS pixel demixing

First, endmembers were selected based on the detection target. In this identification, four categories were selected: water, land, cloud, and ship hull, as shown in Figure 1-4.

Figure 1 Water body end-member

Figure 2 Land end-member

Figure 3 Cloud Meta

Figure 4. Hull end element

The reflectivity of different object types is shown in Figure 5 below.

Figure 5. Hull end element

Results Display

Figure 6 shows the image to be detected, which contains a number of ships in the water.

Figure 6. True-color composite image (Xiguang-1 05 satellite - Tianxianpei 20250715)

The above pixel solution was applied to the data. After removing land areas using mask data and performing threshold classification, the following results were obtained, with red pixels representing the ship's hull:

Figure 7 (Xiguang No. 1 05 satellite - Tianxianpei number 20250715)

Based on FCLS unmixing and subsequent discrimination methods, we counted the number of ship hulls extracted within the study area and compared it with the results of manual interpretation. Overall, the unmixing method can accurately separate areas with high endmember abundance on ship hulls, forming clear ship clustering patterns. For example, in a nearshore port area scenario, the number of ship hulls manually interpreted was 11, while the FCLS-based identification result extracted 13, achieving a recognition rate of approximately 85%. Among them, large ships were almost all correctly identified due to their significant endmember features; however, one small ship was misidentified because its abundance feature was not prominent enough due to its high mixing ratio with the water or interference from the shoreline. Another ship was a small patch with high residuals and was misidentified as a ship, indicating that the method has some false alarms under strong waves or foam interference. Overall, the number of ship hulls extracted by the unmixing results is highly consistent with the manually labeled data, verifying the effectiveness of the method in complex water environments. However, it also suggests the need to further combine spatial morphological constraints and multi-source feature fusion to reduce the false positive and false negative rates.