326
J. Eur. Opt. Society-Rapid Publ. 22, 32( 2026)
image needs to undergo preprocessing such as radiometric calibration, atmospheric correction, and cropping. Then, regions of interest( ROIs) are established to select training samples for each category, and the ROIs for each type of feature are marked with different colors. Figure 3 illustrates the supervised classification process.
The separability between different sample types is determined by calculating the Jeffries-Matusita Distance( JM) and the transformation separation. A maximum likelihood-based classification is selected for sample training to generate a classification model, which is then processed. At last, the accuracy and reliability of the classification are measured by two indicators: the overall accuracy in the confusion matrix and the Kappa coefficient. The JM distance [ 23 ] can be expressed as:
Figure 1. Space-based optoelectronic system detecting targets.
atmospheric transmittance of solar radiation to the background; E sun is the solar irradiance; BRDF b is the bidirectional reflectivity of the background; h i( u i) andh r( u r) are the incident zenith angle( incident azimuth angle) and the reflected zenith angle( reflection azimuth angle), respectively; q b is the background reflectivity; and L sky is the incident radiance of the sky.
2.1 Port temperature distribution inversion
Measured remote sensing images over a coastal region in northern China were selected from Landsat 8 TIRS Band 10 imagery( 30 m). Surface temperature was retrieved using a single-channel algorithm [ 18 ]. Figure 2 illustrates the temperature differences between the sea surface and land in different seasons. In spring and autumn, the temperature difference between the sea and land is small, and the temperature changes are stable. The land surface heats up quickly in summer and cools down quickly in winter. Seawater, due to its high specific heat capacity, experiences smaller temperature fluctuations.
Due to the significant differences in the radiation characteristics of different land cover types and sea surfaces, it is necessary to establish corresponding radiation characteristic models for various land types to improve the physical consistency and accuracy of the inversion results. In addition, due to the limitations of the spatiotemporal resolution of satellite observations, the inversion results only reflect the temperature distribution at a specific transit time. It is difficult to directly characterize the changes in thermal radiation throughout the day or over a continuous period of time. Therefore, it is necessary to extend the temperature time series to obtain multi-temporal temperature fields.
2.2 Port background feature classification
To improve the accuracy of infrared radiation calculation for port backgrounds, a supervised classification method is used to classify the background features. First, the original
� JM ¼ 21�e �B
B ¼ 1 ð 8 m 1 � m 2 Þ 2 2
r 2 1 þ r 2
2 þ 1 2
r2 1 ln þ r2 2
2r 1 r 2 ð3Þ
ð4Þ
where B is the Bhattacharyya distance between the two classes based on a certain feature, m 1 and m 2 are the feature means of the two classes, and σ 1 and σ 2 are the feature standard deviations of the two classes. The calculation results are shown below:
The JM distance range from 0 to 2, where 0 indicates that the two categories are almost completely confused on a certain feature, and 2 indicates that the two categories are completely separated on a certain feature. As shown in the Table 1, the separability values are all greater than 1.9, indicating reasonable sample separation. The images are classified according to the maximum likelihood method, and the results are shown in Figure 4.
As shown in Figure 4, the port background features can be roughly divided into four categories, with the following percentages: water 81.54 %, vegetation 2.92 %, land 7.45 %, and buildings 8.09 %. A confusion matrix is used to evaluate the classification results, determining the accuracy and reliability of the classification. Two indicators, Overall Accuracy( OA) and Kappa Coefficient [ 24 ], are used for evaluation:
P i; j
Kappa ¼ 1 � w i; jQ i; j
Pi; j w ð5Þ i; jE i; j w i; j ¼ ði � jÞ2 ð6Þ ðN � jÞ 2
where w i, j is the squared weighted average; O i, j is the element in the observation matrix; E i, j is the element in the expectation matrix; N is the total number of classifications. The classification verification results are follows, the overall accuracy is 0.9984, which is higher than 85 %. Kappa