Change detection (CD) is an important application of remote sensing, which provides information about land cover changes on the earth's surface. Hyperspectral image (HSI) can show more spectral information, which greatly improves the ability of remote sensing to identify change features. The challenge is how to overcome the scarcity of labeled samples and extract the change information of high-dimensional spectra in HSI. To solve the previous problem, a semi-supervised CD with multi-layer cascade screening strategy (MCS4CD) that uses both the spatial information and active learning is proposed to select highly reliable unlabeled samples to increase the training sets. The MCS4CD method can efectively use unlabeled samples to improve accuracy. Additionally, a subspace CD method based on iterative slow feature analysis (ISFA) and principal component analysis (PCA) is designed to extract the most temporally invariant component from the high dimensional space. Experimental results on a hyperspectral dataset show that with a small umber of labeled samples, the proposed method achieves a much better performance than existing CD methods. Change detection, hyperspectral image (HSI), iterative slow feature analysis (ISFA), semi-supervised learning, active learning The surface ecosystem and human social activities are dy- classifies data samples based on the features inherent
1. Introduction
namically developing and evolving[
Over the past decades, land-use and land-cover CD
creasing attention in the supervised, unsupervised, and
semi-supervised algorithms. Supervised learning (SL)
nEvelop-O
LGOBE
CDCEO 2021: 1st Workshop on Complex Data Challenges in Earth
vector analysis (CVA)[
Step 1): Circular neighborhood (CN): We empirically
Based on the above analysis of the current hyperspec- find out that 4 or 8 neighborhood are usually used to tral CD problems, it is obvious that we need to explore obtain the SNI. It only covers a small area within a fixed CD algorithms by focusing on two main points. Firstly, radius, which obviously cannot meet the needs of diferit is often dificult for ISFA algorithm to separate the ent sizes. 4 or 8 neighborhood windows are too small changed and unchanged pixels in HSI classification espefor searching useful unlabeled samples. In this paper, cially with limited small training samples. In this paper, we adopt a CN window, which can adjust the search a novel semi-supervised classification algorithm based on multi-layer cascade screening strategy (MCS4CD) is put forward. In the semi-supervised process, the spatial radius d to find the optimal size (including 4- and 8neighborhood).
Step 2): SNI extraction strategy: “Tobler’s First Law neighborhood of labeled training samples is combined of Geography” gives us an important assumption that with active learning (AL) to select the most helpful unlathe label category of unlabeled samples should be conbeled samples[16], which is used as the pseudo labeled sistent with the existing training sample categories in set to retrain the support vector machine (SVM) classifier. the spatial neighborhood area. However, in the process Secondly, the performance of ISFA algorithm is degraded of determining unlabeled samples, the positive efect of due to the band redundancy of HSI. To solve this problem, SNI on SSL method is often ignored. Based on the iniwe designed a new CD algorithm using PCA and ISFA.
tial classification results and SNI, the labels of unlabeled The main contributions of this paper are summarized
samples are screened for the second time. The second 1) The reliability of selected unlabeled samples is in- quality and wrong labels.
The proposed MCS4CD strategy combines spatial neigh- samples with high confidence or most informative are as follows: creased with the proposed MCS4CD strategy that utilizes the spatial information and AL algorithm.
2) The MCS4CD strategy takes into account the positive efect of neighborhood spatial information in the semi-supervised classification process.
3) The PCA+ISFA method is designed for extracting the unchanged features of bi-temporal HSI data.
borhood information (SNI) extraction strategy with AL algorithm to select the most informative unlabeled samples as the pseudo labeled set to further improve classification performance. The specific procedures of MCS 4CD are shown in Figure 1.
If the label categories of unlabeled samples are determined only by the primary SVM classification map, it is dificult to ensure satisfactory accuracy. This is because the primary classification accuracy is not high, the labels of the candidate set are misclassified. Therefore, the subsequent SSL process will be afected by the error labels, resulting in error accumulation. In this paper, MCS4CD strategy is constructed to help the SVM classifier label the selected unlabeled samples. screening strategy eliminates the negative efects of low
In the SSL process, a large number of unlabeled data points are selected as the candidate sets. To further simplify the samples, AL strategy is used to select the most informative unlabeled samples from the candidate sets. AL aims to carefully choose the samples to be labeled to achieve a higher accuracy while using as few requests as possible, thereby minimizing the cost of obtaining labeled data. After the third screening step, the unlabeled selected as the final pseudo samples through AL.
In this paper, breaking ties (BT) query strategy of AL algorithm is used to collect the most informative unlabeled samples. The decision criterion of BT is: ̂ = arg min {max ( = | ) − max ( = | )}} ,∈
∈ ( = | ) = ∈ \{ 1
+} 1 + exp( ( ) + ) (1) (2)
Where + = arg max∈ ( = | ) represents the class label corresponding to the largest posterior probability for sample , and ∈ \{ class labels excluding +. is provided the probabilistic outputs by the probability model-based SVM. +} represents the interested
To increase the reliability of selected unlabeled samples, a SSL method that is based on the MCS4CD strategy is adopted. The detailed strategy is described as follows:
Step 1) Initialize training samples = {( 1, 1), ...( 1, )}, and set parameters: spectral dimension of subspace feature , radius d of CN, the number of training samples for each class p;
Step 2) Extract subspace spectral feature ∈ ℝ using the PCA method;
Step 3) Extract changed and unchanged features ∈ ℝ using the ISFA method;
Step 4) Train SVM probability model to predict the label ̂ = { ̂ , = 1, ..., } of unlabeled samples;
Step 5) Select a circle neighborhood which takes the selected as the center and remove redundancy samples (including background information and repeated selection of training samples). Retain samples with the same labels as spatial neighborhood label. The extracted candidate samples are denoted as = {( 1, 1), ...( 1, )}, = 1, ..., ;
Step 6) Simplify to by using the AL algorithm. The is expanded with each iteration of BT. Then, update the labeled sample sets + = { , };
Step 7) Test the performance of the final pseudo labeled samples using the SVM classifier.
The USA Dataset illustrates an irrigated agricultural field of Hermiston city in Umatilla County, Oregon, OR, the USA, which was collected on May 1, 2004, and May 8, 2007, respectively. The size of this dataset is 307 lines by 241 samples, with 154 spectral bands. The land cover types include soil, irrigated fields, river, cultivated land and grassland. For this dataset, all changes related to the type of land cover and river. The true color composite image of the USA dataset and its corresponding landcover CD map are shown in Figure. 2.
To demonstrate the efectiveness of the proposed method, we will display the numerical results on the USA dataset, as shown in Figure 3. As can be observed in table 1, MCS4CD is capable of building better classification performance that compensates for the lack of labeled training data. MCS4CD projects the original data into a new transformed space to better separate the changed and unchanged pixels.
For the MCS4CD, one of the key questions is how to confirm the suitable radius d, which influences the accuracy and the numbers of selected unlabeled samples. Figure. 4 shows the result of USA data. When the number of pseudo labeled samples is 0, 20, 40, 60, 80,100, 120 and the search radius d ranges from 1 to 3, the results corresponding d =2 is the best.
This study proposed a semi-supervised CD method with slow feature analysis, including multi-layer cascade screening strategy and data transformation strategy in spectral subspace domain, for HSI classification with a small number of labeled samples. In the semi-supervised process, the slow feature extraction in high-dimensional space, the use of BT algorithm, and the decision strategy for the label of unlabeled samples are all key points. On the one hand, we use PCA method to reduce too many spectral bands, which has a negative impact on the expected CD performance. On the other hand, BT, circular neighbor and SVM are combined together to improve the judgment accuracy of unlabeled samples. Experimental results with HSI indicate that the proposed MCS4CD approach can obtain well performance. (a) CVA (b) MAD (c) IRMAD (d) ISFA (e) DPCA (f) PCA+ISFA (g) MCS4CD (h) Ground truth use change detection and analysis using multitemporal and multisensor satellite data, International Journal of Remote Sensing 29 (2008) 4823–4838. [15] C. Wu, B. Du, L. Zhang, Slow feature analysis for change detection in multispectral imagery, IEEE Transactions on Geoscience and Remote Sensing 52 (2013) 2858–2874. [16] C. Liu, J. Li, L. He, Superpixel-based semisupervised active learning for hyperspectral image classification, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12 (2018) 357–370.