Author Affiliations
1 School of Aeronautics and Astronautics, Shanghai Jiao Tong University, Shanghai 200240, China2 Room 15, Institute of Shanghai Satellite Engineering, Shanghai 201108, China3 Satellite Remote Sensing Application Technology Laboratory, Shanghai Institute of Meteorological Science, Shanghai 200030, Chinashow less
Fig. 1. Diagrams of hyperspectral data and site location of Suzhou. (a) Diagram of station location; (b) February 28, 2015, passage 150; (c) March 18, 2015, passage 179
Fig. 2. Average spectral curves at underlying surface of Suzhou under different haze conditions
Fig. 3. Random sampling spectral curves of haze (red) and non-haze (blue)
Fig. 4. Scatter plots of haze (red) and non-haze (blue) with PCA characteristics
Fig. 5. Schematic of residual learning
Fig. 6. Framework of deep residual network for hyperspectral haze monitoring
Fig. 7. Schematic of internal structure of residual block
Fig. 8. Comparison of performance of CNN and ResNet with different network depths
Fig. 9. Structure analysis of DBN
Fig. 10. Comparison of network performance with large training sampling. (a) Training error; (b) test error
Fig. 11. Experiment results of large training samples of BP, CNN-13 and ResNet-13
Fig. 12. Haze monitoring results of Suzhou on January 26, 2015. (a) Diagram of site location; (b) result of SVM; (c) result of DBN; (d) result of Resnet
Fig. 13. Results of Shanghai haze monitoring on January 4, 2015. (a) Diagram of site location; (b) result of SVM; (c) result of DBN; (d) result of Resnet
Number of experiments | SVM | BP | DBN | CNNs | ResNet |
---|
OA | Kappa | | OA | Kappa | | OA | Kappa | | OA | Kappa | | OA | Kappa |
---|
1 | 0.9398 | 0.9071 | 0.9388 | 0.9053 | 0.9492 | 0.9217 | 0.9473 | 0.9194 | 0.9608 | 0.9398 | 2 | 0.9419 | 0.9109 | 0.9405 | 0.9081 | 0.9498 | 0.9229 | 0.9463 | 0.9181 | 0.9671 | 0.9495 | 3 | 0.9474 | 0.9189 | 0.9484 | 0.9202 | 0.9512 | 0.9248 | 0.9500 | 0.9233 | 0.9619 | 0.9414 | Average | 0.9430 | 0.9123 | 0.9426 | 0.9112 | 0.9501 | 0.9231 | 0.9479 | 0.9203 | 0.9633 | 0.9436 |
|
Table 1. Experimental results of haze recognition with different methods
Term | ResNet result | CNNs result |
---|
Non | Mild | Moderate | Severe | | Non | Mild | Moderate | Severe |
---|
Non | 249617 | 2148 | 1027 | 43 | 240634 | 2520 | 1416 | 36 | Mild | 3726 | 120769 | 2536 | 0 | 8228 | 120416 | 3481 | 0 | Moderate | 3919 | 3241 | 92555 | 0 | 7439 | 3223 | 91224 | 0 | Severe | 203 | 1 | 15 | 32867 | 1155 | 0 | 12 | 32874 | Total | 257465 | 126159 | 96133 | 32910 | 257456 | 126159 | 96133 | 32910 | Classification accuracy /% | 96.95 | 95.73 | 96.28 | 99.87 | 93.47 | 95.45 | 94.89 | 99.89 | Overall accuracy /% | 96.71 | 94.63 |
|
Table 2. Confusion matrix of haze classification