≪ View Page

Diferenzas

Isto amosa as diferenzas entre a revisión seleccionada e a versión actual da páxina.

Enlazar a esta vista de comparación

Both sides previous revisionPrevious revision
Next revision		Previous revision
hiperespectral:sae-cd [2018/01/16 17:18] – [Downloads] javier.lopez.fandinohiperespectral:sae-cd [2018/04/03 11:00] (actual) – [STACKED AUTOENCODERS FOR MULTICLASS CHANGE DETECTION IN HYPERSPECTRAL IMAGES] javier.lopez.fandino
Liña 1: Liña 1:
-Experimental results and addictional information related to the paper "STACKED AUTOENCODERS FOR MULTICLASS CHANGE DETECTION IN HYPERSPECTRAL IMAGES".+====== STACKED AUTOENCODERS FOR MULTICLASS CHANGE DETECTION IN HYPERSPECTRAL IMAGES ====== 
 + 
 +Experimental results and addictional information related to the paper "STACKED AUTOENCODERS FOR MULTICLASS CHANGE DETECTION IN HYPERSPECTRAL IMAGES", accepted in the International Geoscience and Remote Sensing Symposium, IGARSS 2018.
  
 ==== Abstract ==== ==== Abstract ====
  
-Change detection (CD) in multitemporal datasets is a key task in remote sensing. In this paper, a scheme to perform multiclass CD for remote sensing hyperspectral datasets extracting features by means of Stacked Autoencoders (SAEs) is introduced. The scheme combines multiclass and binary CD to obtain an accurate multiclass change map. The multiclass +Change detection (CD) in multitemporal datasets is a key task in remote sensing. In this paper, a scheme to perform multiclass CD for remote sensing hyperspectral datasets extracting features by means of Stacked Autoencoders (SAEs) is introduced. The scheme combines multiclass and binary CD to obtain an accurate multiclass change map. The multiclass CD begins with the fusion of the multitemporal data followed by feature extraction by SAE. The binary CD is based on the spectral nformation by calculating pixel-wise distances and thresholding, and it also incorporates spatial information through watershed segmentation. The data coming from the multiclass CD is filtered by using the binary CD map and later classified by a Support Vector Machine or an Extreme Learning Machine algorithm. The scheme was evaluated over a multitemporal hyperspectral dataset obtained from the Hyperion sensor. Experimental results show the effectiveness of the proposed scheme using SAE for extracting the relevant features of the fused information when compared to other published feature extraction methods
-CD begins with the fusion of the multitemporal data followed by feature extraction by SAE. The binary CD is based on +
-the spectral information by calculating pixel-wise distances and thresholding, and it also incorporates spatial information through watershed segmentation. The data coming from the multiclass CD is filtered by using the binary CD map and later classified by a Support Vector Machine or an Extreme Learning Machine algorithm. The scheme was evaluated over a multitemporal hyperspectral dataset obtained from the Hyperion sensor. Experimental results show the effectiveness of the proposed scheme using SAE for extracting the relevant features of the fused information when compared to other pub- +
-lished feature extraction methods+
  
  
-===== Downloads ===== +==== Input datasets ====
- +
-=== Input datasets ===+
  
 //All the images are avaiable in Matlab (.mat) format, among others. For further information see the readme in the files.// //All the images are avaiable in Matlab (.mat) format, among others. For further information see the readme in the files.//
Liña 17: Liña 14:
 * [[https://citius.usc.es/investigacion/datasets/hyperspectral-change-detection-dataset|Hermiston]]  * [[https://citius.usc.es/investigacion/datasets/hyperspectral-change-detection-dataset|Hermiston]] 
  
-=== Results === +==== Experimental setup ====
- +
-== Experimental setup ==+
  
 * Codes were run in Ubuntu 14.04.  * Codes were run in Ubuntu 14.04. 
 +
 * Caffe framework 1.0.0-rc3 to perform the feature extraction by means of SAE. * Caffe framework 1.0.0-rc3 to perform the feature extraction by means of SAE.
-* The SAE is configured to obtain 12 features. +  * The SAE is configured to obtain 12 features. 
-* Two consecutive layers reduce the dimensionality of the data from 242 to 100 and from 100 to 12 features,respectively. +  * Two consecutive layers reduce the dimensionality of the data from 242 to 100 and from 100 to 12 features,respectively. 
-* The SAE is trained with 20% of the available pixels randomly chosen. +  * The SAE is trained with 20% of the available pixels randomly chosen. 
-* A batch of 64 pixels per iteration is used +  * A batch of 64 pixels per iteration is used. 
-* The iteration limit is fixed to 300000 iterations. +  * The iteration limit is fixed to 300000 iterations. 
-* The back-propagation process uses a Stochastic Gradient Descent (SGD) and the ’inv’ learning rate policy [inv = base lr ∗ (1 + γ ∗ i)^(−power)] being i the iteration number and with a base learning rate (base lr) of 0.01, and values for the parameters γ and power of 0.0001 and 0.75 respectively. +  * The back-propagation process uses a Stochastic Gradient Descent (SGD) and the ’inv’ learning rate policy [inv = base lr ∗ (1 + γ ∗ i)^(−power)] being i the iteration number and with a base learning rate (base lr) of 0.01, and values for the parameters γ and power of 0.0001 and 0.75 respectively. 
  
 * NWFE and PCA used for comparision purposes retaining 12 features. * NWFE and PCA used for comparision purposes retaining 12 features.
  
 * ELM and SVM trained with 5% of the reference data available for each class. * ELM and SVM trained with 5% of the reference data available for each class.
-* Training samples randomly chosen in each run.  +  * Training samples randomly chosen in each run.  
-* 10 independent runs for each classifier. +  * 10 independent runs for each classifier. 
-* SVM classification carried out using the LIB-SVM library and the Gaussian radial basis function (RBF)  +  * SVM classification carried out using the LIB-SVM library and the Gaussian radial basis function (RBF).  
-* ELM configured with a sigmoidal activation function.+  * ELM configured with a sigmoidal activation function. 
 + 
 +==== Outputs ==== 
 + 
 +=== Image files === 
 +|Reference data of changes |Binary CD map |Multiclass CD map| 
 +|{{:hiperespectral:referencedatacolorhermiston5.png?200|}}|{{:hiperespectral:binarycd.png?200|}}|{{:hiperespectral:svmcolorhermiston.png?200|}}| 
 + 
 + 
 + 
 +=== Accuracy results === 
 +==Binary CD accuracies== 
 +|**Corect** |**Missed Alarms**|**False Alarms** |**Total Error**| 
 +|77020 (98.74%) |509 |471 |980 (1.25%) | 
 + 
  
 +==Multiclass CD accuracies==
 +|**Classifier**   | **Parameters**     |**FE**   | **OA (%)**  | **AA (%)**  | **Kappa**  |
 +| ELM             | N=120              | PCA     | 91.73       | 76.06       | 86.83      |
 +| ELM             | N=120              | NWFE    | 91.76       | 76.75       | 86.83      |
 +| ELM             | N=60               | SAE     | 95.19       | 90.45       | 92.31      |
 +| SVM             | C: 64.0 γ: 32.0    | PCA     | 91.46       | 71.16       | 86.46      |
 +| SVM             | C: 32.0 γ: 16.0    | NWFE    | 91.29       | 90.61       | 86.05      |
 +| SVM             | C: 32.0 γ: 0.0625  | SAE     | 95.52       | 92.56       | 92.90      |
  
 +C: penalty term in the training of the SVM. γ: radius of the gaussian function of the SVM. N: Number of neurons in the hidden layer of the ELM. FE: Feature Extraction method.
 ===== License ===== ===== License =====
  
 :cc-by-nc-nd:   :cc-by-nc-nd: