Joint Solution of Urban Structure Detection from Hyperion Hyperspectral data Lin Cong, Brian Nutter, Daan Liang Wind Science and Engineering Department of Electrical and Computer Engineering Department of Construction Engineering & Engineering Technology Texas Tech University, Lubbock, TX, USA e-mail: {lin.cong, brian.nutter, daan.liang}@ttu.edu 7000 INTRODUCTION directional energy in Fourier domain direction of maximum energy Problem: Damage condition of residential areas are more concerned than that of natural areas in post-hurricane damage assessment. Recognition of residential and natural areas from commonly used low-spatial-resolution hyperspectral images is thus important. (a) directional energy in Fourier domain 6000 (b) 10 gray level j 20 30 5000 4000 3000 2000 40 1000 50 Solution: A spatial-feature extraction method based on hierarchical Fourier transform – Co-occurrence matrix is developed. Spatial and spectral features are then combined to a joint feature vector. Best feature combinations are determined by K-fold cross validation. 60 0 10 20 30 40 gray level i 50 60 0 20 40 60 80 100 angle (degree) 120 140 160 180 (c) (d) Figure 2. (a) A sample of residential region; (b) The Fourier transform of the residential region; (c) Plot of the directional energy distribution in Fourier domain; (d) Co-occurrence matrix calculated with the direction of the maximum energy and offset equal to one; x 10 2.5 12 directional energy in Fourier domain direction of the maximum energy METHOD (b) (a) Flow chart directional energy in Fourier domain 2 10 20 gray level j Hyperspectral data 30 1.5 Spectral correlation Fourier Transform Co-occurrence matrix PCA components 60 0 10 20 30 40 gray level i 50 60 20 40 60 80 100 angle (digree) 120 140 160 180 (c) Figure 3. (a) Another sample of residential region; (b) The Fourier transform of the residential region; (c) Plot of the directional energy distribution in Fourier domain; (d) Co-occurrence matrix calculated with the direction of the maximum energy and offset equal to one; 2 x 10 11 directional energy in Fourier domain direction of the maximu energy 1.8 Feature selection K-means clustering directional energy in Fourier domain 1.6 (b) (a) 10 20 gray level j Bayes Classification 0 (d) Texture measures 30 50 Datasets 1.4 1.2 1 0.8 0.6 40 Error 1.88% 1.90% 1.97% 2.01% 2.02% 2.04% 2.05% 2.07% 2.08% 2.10% 2.10% 2.11% 2.13% 2.13% 2.13% 2.15% 2.15% 2.16% 2.16% 2.17% Rank Combination 21 1111011000 22 1101100000 23 1010111001 24 1011001001 25 1011011001 26 1111101001 27 1011011000 28 1011010001 29 1111101000 30 1111110001 31 1001011001 32 1111110000 33 1111001000 34 1111100001 35 1111010000 36 1101101001 37 1111111010 38 1101010000 39 1111111011 40 1111100000 Error 2.17% 2.18% 2.19% 2.21% 2.21% 2.21% 2.23% 2.23% 2.24% 2.25% 2.26% 2.26% 2.27% 2.28% 2.28% 2.29% 2.30% 2.31% 2.31% 2.31% Rank 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 Combination 0000100000 0010000110 0010000111 0000000111 0010100110 0000000110 0000100110 0010100101 0010000101 0000100101 0000000101 0011000010 0101000000 0001000010 0010000010 0000000010 0011000000 0001000000 0010000000 0100000000 Error 12.5% 12.7% 12.7% 12.7% 12.8% 12.8% 12.9% 13.0% 13.0% 13.1% 13.2% 13.2% 13.7% 14.7% 15.3% 15.8% 16.1% 19.4% 24.4% 27.2% Spectral Solution (avg. error: 17.39%) Spatial Solution (avg. error: 19.34%) Joint Solution (avg. error: 12.99%) Rank Combination 1 1110001001 2 1110011001 3 1110111000 4 1110001000 5 1110010001 6 1110101000 7 1110011000 8 1110010000 9 1110110000 10 1110101001 11 1110110001 12 1110111001 13 1111011000 14 1111001000 15 1111001001 16 1111011001 17 1111010000 18 1111111000 19 1111010001 20 1110100001 Error 5.67% 5.77% 5.81% 5.84% 5.84% 5.86% 5.87% 5.89% 5.96% 6.01% 6.03% 6.06% 6.19% 6.29% 6.29% 6.34% 6.39% 6.42% 6.50% 6.57% Rank Combination 21 1111111001 22 1111101000 23 1111110001 24 1111101001 25 1110000001 26 1111100001 27 1111110000 28 1111100000 29 1111000001 30 1110011010 31 1110101010 32 0110111001 33 1110110010 34 1110001100 35 1110111011 36 1110101011 37 1110111100 38 0110011001 39 1110111010 40 0110001001 60 20 30 40 gray level i 50 0.2 60 0 20 40 60 80 100 angle (degree) 120 140 160 Error 6.60% 6.63% 6.66% 6.68% 6.75% 6.80% 6.83% 6.93% 7.12% 7.36% 7.39% 7.43% 7.44% 7.50% 7.52% 7.55% 7.56% 7.58% 7.58% 7.59% Rank 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 Combination 0000100010 1101000010 1000100010 0101000010 0001000010 1000000010 1000100000 1001000010 1011000000 0011000000 0100000000 0101000000 0000100000 1010000000 1100000000 0001000000 1101000000 0010000000 1001000000 1000000000 Error 18.4% 18.5% 18.7% 19.1% 21.0% 21.6% 21.7% 22.2% 24.0% 24.6% 26.4% 26.5% 26.6% 27.0% 27.0% 27.7% 28.1% 28.9% 31.8% 39.8% Results 180 (c) (d) Classified as Natural or River 3106 25617 Error Rate Classified as Residential 4.59% Residential 62704 38.28% Natural + River 16116 Classified as Natural or River 5011 25389 Error Rate Classified as Residential 7.40% Residential 65225 38.83% Natural + River 11699 Classified as Natural or River 2490 29806 Error Rate 3.68% 28.19% 1. Training data of New Orleans dataset is used to train the Bayes classifier, and then the Lubbock dataset is classified. Figure 9. “Cross” classification results of Lubbock dataset. Blue: residential areas correctly classified; Green: natural areas correctly classified; Red: residential areas misclassifed as natural areas; Pink: natural areas misclassifed as residential areas. (a) Purely spectral (b) Purely spatial Spectral Solution (avg. error: 69.95%) Spatial Solution (avg. error: 12.87%) Joint Solution (avg. error: 20.25%) Classified as Residential Classified as Natural Error Rate Residential Region 40214 18451 31.35% Residential Region 55192 3473 5.92% Classified as Residential Classified as Natural Error Rate Residential Region 56524 2141 3.65% Classified as Residential Classified as Natural Error Rate (5) Angular Second Moment (ASM) ASM Pi , j 2 2 i , j 1 i , j 1 (2) Dissimilarity (DIS) (6) Maximum Probability (MAX) N DIS Pi , j i j MAX max( Pi , j ) i , j 1 (a) (b) (c) (d) (3) Homogeneity (HOM) HOM (7) Entropy (ENT) N ENT Pi , j log 2 Pi , j Pi , j N 1 (i j ) i , j 1 2 (a) Ground truth i , j 1 (4) Similarity (SIM) SIM (b) Purely spectral (c) Purely spatial (d) Joint solution Figure 7. Results of Bayes classification for Lubbock dataset. (a) Manually made ground truth; (b) – (d) Results by using purely spectral features, purely spatial features, joint features, respectively. Blue: residential areas correctly classified; Green: natural areas correctly classified; Red: residential areas misclassifed as natural areas; Pink: natural areas misclassifed as residential areas. Pi , j N 1 i j i , j 1 Table 3: Error rates of the Bayes classification for Lubbock dataset Spectral Solution (avg. error: 15.45%) Spatial Solution (avg. error: 13.43%) CON (e) (f) (g) (h) DIS HOM SIM ASM MAX ENT Figure 5. Texture measures of Lubbock dataset Joint Solution (avg. error: 10.84%) Classified as Residential Classified as Natural Error Rate Residential Region 50443 8222 14.20% Residential Region 48319 10346 21.26% Classified as Residential Classified as Natural Error Rate Residential Region 51127 7538 12.85% Classified as Residential Classified as Natural Error Rate Natural Region 12281 61734 16.59% Natural Region 7479 66536 10.10% Natural Region 6848 67168 9.25% Figure 1. (a) Original hyperspectral image taken over Lubbock, TX in 01/2003; (b) – (c) The top two significant PCA bands of Lubbock dataset; (d) Spectral correlation against the spectrum of construction asphalt; (e) Original hyperspectral image taken over New Orleans, LA in 04/2005; (f) – (g) The top two significant PCA bands of New Orleans dataset; (h) Spectral correlation against the spectrum of construction asphalt; Fourier transform – Co-occurrence matrix • Residential areas display periodic street patterns while the natural areas are universal. • Fourier Transform is applied to detect the directions orthogonal to the street patterns. • Gray level co-occurrence matrix is calculated between neighboring pixels with an offset of one in the direction orthogonal to the street patterns. CON DIS 18.38% Natural Region 24721 49294 33.40% Figure 10. “Cross” classification results of New Orleans dataset. Blue: residential areas correctly classified; Green: natural areas correctly classified; Red: residential areas misclassifed as natural areas; Pink: natural areas misclassifed as residential areas. N N Natural Region 69053 4962 93.30% Natural Region 13607 60408 2. Training data of Lubbock dataset is used to train the Bayes classifier, and then the New Orleans dataset is classified. Texture measures (1) Contrast (CON) (c) Joint solution Table 5: Error rates of the “cross” classification for Lubbock dataset Bayes Classification Figure 4. (a) A sample of natural region; (b) The Fourier transform of the natural region; (c) Plot of the directional energy distribution in Fourier domain; (d) Co-occurrence matrix calculated with the direction of the maximum energy and offset equal to one; CON Pi , j (i j ) Natural + River 15888 “Cross” Bayes Classification 0.4 10 Classified as Residential Residential 64609 Table 2: Performance of a subset of all joint feature combinations for New Orleans dataset. 50 PCA transform Rank Combination 1 1100111001 2 1111111001 3 1101011001 4 1101011000 5 1101001001 6 1101010001 7 1110111001 8 1101101000 9 1111001001 10 1111011001 11 1011100000 12 1101110000 13 1001100000 14 1100111000 15 1111111000 16 1101111000 17 1111010001 18 1110111000 19 1101111001 20 1101100001 Table 4: Error rates of the Bayes classification for New Orleans dataset 1 0.5 40 Table 1: Performance of a subset of all joint feature combinations for Lubbock dataset. Features are listed in the combinations following the order: PCA1, PCA2, spectral correlation, CON, DIS, HOM, SIM, ASM, MAX, ENT. A “1” means that the feature in the associated position is selected in the combination, and a “0” means that associated feature is not selected. HOM SIM ASM MAX (a) Purely spectral (b) Purely spatial (c) Joint solution Table 6: Error rates of the “cross” classification for New Orleans dataset Spectral Solution (avg. error: 42.07%) Spatial Solution (avg. error: 18.20%) Joint Solution (avg. error: 18.20%) Classified as Residential Classified as Natural or River Error Rate Classified as Residential Classified as Natural or River Error Rate Classified as Residential Classified as Natural or River Error Rate Residential 61881 5834 8.62% Residential 53840 13875 20.49% Residential 53937 13778 20.35% Natural + River 40110 1395 96.64% Natural + River 5999 35506 14.45% Natural + River 6079 35426 14.65% Conclusion 1. Improved accuracy in Bayes classification between residential and natural areas was achieved by using both spectral and macroscopic spatial information. 2. The spatial features extracted by proposed Fourier transform – Co-occurrence matrix method seem to be reliable in “cross” classification, although the purely spectral information between different datasets is so different that it fails the cross classification. ENT Figure 6. Texture measures of New Orleans dataset Future work Feature Selection K-fold cross validation is applied on the training dataset to determine the best combinations of the spectral and spatial features. This material is based upon work supported by the National Science Foundation under Grant No. 0800487 (a) Ground truth (b) Purely spectral (c) Purely spatial (d) Joint solution Figure 8. Results of Bayes classification for New Orleans dataset. (a) Manually made ground truth; (b) – (d) Results by using purely spectral features, purely spatial features, joint features, respectively. Blue: residential areas correctly classified; Green: natural areas correctly classified; Red: residential areas misclassifed as natural areas; Pink: natural areas misclassifed as residential areas. 1. More testing and verification on additional datasets are needed in the future. 2. The segmentations of residential and natural areas can be used for model choice in spectral unmixing. 3. The spectral unmixing results at the same position before and after a hurricane can be compared to assess the damage level.