合成孔径雷达（Synthetic Aperture Radar，SAR）的微波成像技术具有全天时、全天候的特性，在民用和军用领域得到广泛应用。SAR图像变化检测是指对同一地点两幅不同时相的SAR图像通过合适的算法得到变化区域的过程，该技术在环境监测、农田检测、自然灾害检测等方面的应用日益重要。然而，SAR图像存在大量相干斑噪声，对变化检测产生很大影响。当前，提高SAR图像变化检测准确率仍是关键目标。本文以SAR图像为研究对象，深度森林网络为基本方法，着重针对SAR图像中的相干斑噪声对变化检测产生的影响展开研究。

     （1）针对SAR图像中的相干斑噪声问题，本文提出了一种基于深度森林网络的特征融合（简称RF-MGS-DF）与SAR图像变化检测方法。首先，提取了差异图两种不同尺度的邻域特征，通过深度森林（deep forest，DF）网络的多粒度扫描（multi granularity scanning，MGS）结构将两种不同尺度的邻域特征转化为类概率特征；然后，提取了差异图的纹理特征，通过随机森林（random forest，RF）特征重要性评分去除冗余特征；最后，将类概率特征与筛选后的纹理特征进行串联融合，采用DF进行决策分类。实验结果表明，邻域特征经过MGS后假阳性有所下降，抑制噪声性能更好；纹理特征经过RF特征筛选后，假阴性有所下降，检测精度更高；RF-MGS-DF方法得到的变化检测结果与特征融合前相比噪声减少，变化边缘更加细致，总体精度和Kappa系数有所上升。

     （2）针对DF网络在SAR图像变化检测上的局限性，本文提出了一种基于异构深度森林网络的特征筛选（简称RFHDF）与SAR图像变化检测方法。首先，为了增大网络的多样性，提高泛化性能，在DF的基础上改进得到异构深度森林网络（heterogeneous deep forest，HDF），该网络每层由RF、完全随机树（extremely randomized trees，ERT）、梯度提升决策树（gradient boosting decision tree，GBDT）、完全梯度提升决策树（extreme gradient boosting，XGBoost）、轻梯度提升机（light gradient boosting machine，LightGBM）五种基本模型组成；其次，为了充分利用HDF每层输出的特征，在HDF每层之间添加一个RF特征筛选模块，将每层输出的重要特征保留下来，作为下一层的输入特征。经过消融实验和对比试验发现，RFHDF方法所得到的变化检测结果噪声变少，准确率和Kappa系数有所提高。

The microwave imaging technology of Synthetic Aperture Radar (SAR) possesses all-weather and all-day capabilities, and is widely utilized in civilian and military sectors. The SAR image change detection refers to the process of identifying changing regions by analyzing two SAR images of the same location acquired at different times using appropriate algorithms. This technology is becoming increasingly important in applications such as environmental monitoring, agricultural land detection, and natural disaster detection. However, SAR images are plagued by a significant amount of speckle noise, which has a substantial impact on change detection. Therefore, enhancing the accuracy of SAR image change detection remains a key objective. This article focuses on SAR images as the research subject, with deep forest networks as the fundamental method, specifically investigating the impact of coherent speckle noise in SAR images on change detection.

(1) In response to the incoherent speckle noise issue in SAR images, this article proposes a feature fusion method based on deep forest networks (referred to RF-MGS-DF) for SAR image change detection. Firstly, two different scale neighborhood features of the difference map are extracted, and converted into probabilistic features by multi-granularity scanning (MGS) structure of deep forest (DF) network. Then, the texture features of the difference map were extracted, and the redundant features were removed by random forest (RF) feature importance score. Finally, the class probability features and the selected texture features are fused in series, and DF is used for decision classification. The experimental results indicate that after undergoing MGS, false positives of neighborhood features are reduced, leading to improved noise suppression performance. Texture features, post RF feature selection, showed a decrease

in false negatives, enhancing detection accuracy. The change detection results from the RF-MGS-DF method exhibited reduced noise, finer change edges compared to before feature fusion, and an increase in overall accuracy and Kappa coefficient.

(2) To address the limitations of DF networks in SAR image change detection, this paper proposes a feature selection method based on heterogeneous deep forest networks (referred to as RFHDF) for SAR image change detection. To increase the diversity of the network, an improvement over the DF called the Heterogeneous Deep Forest (HDF) was developed. This network consists of RF, Extremely Randomized Trees (ERT), Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) at each layer. This enhancement aims to boost the feature transformation capabilities of the network. To fully utilize the features output by each layer of HDF, an RF feature selection module is added between each layer of HDF to retain the important features from the output of each layer and use them as input features for the next layer. Through ablation and comparative experiments, it was found that the change detection results obtained by the RFHDF method had less noise, and the accuracy and Kappa coefficient were improved.

[1] Chugtai A H, Abbasi H, Karas I R. A review on change detection method and accuracy assessment for land cover[J]. Remote Sensing Applications: Society and Environment, 2021, 22:100482

[2] Kang A R, Byun Y G, Chae T B. Development and Evaluation of a Texture-Based Urban Change Detection Method Using Very High Resolution SAR Imagery[J]. Korean Journal of Remote Sensing, 2015, 31(3):255-265.

[3] Ghara F M, Shokouhi S B, Akbarizadeh G. A New Technique for Segmentation of the Oil Spills From Synthetic-Aperture Radar Images Using Convolutional Neural Network[J]. IEEE Applied Earth Observations and Remote Sensing, 2022, vol. 15, 8834-8844.

[4] Kuang X F, Guo J, Wang H Y, et al. Agricultural Field Boundary Delineation Using a Cascaded Deep Network Model From Polarized SAR and Multispectral Images[J]. IEEE Applied Earth Observations and Remote Sensing, 2023 vol. 16, 7228-7247.

[5] Cui J, Jia H, Wang H, et al. A Fast Threshold Neural Network for Ship Detection in Large-Scene SAR Images[J]. IEEE Applied Earth Observations and Remote Sensing, 2022 vol. 15, 6016-603.

[6] Zheng Z, Zhong Y F, Wang J J, et.al. Building damage assessment for rapid disaster response with a deep object-based semantic change detection framework: From natural disasters to man-made disasters[J]. Remote Sensing of Environment, 2021, 265:112636.

[7] Le T T, Froger J L, Minh D H. Multiscale framework for rapid change analysis from SAR image time series: Case study of flood monitoring in the central coast regions of Vietnam[J]. Remote Sensing of Environment, 2022, 269:112837.

[8] Anjali G, Deepak S, Harani M et.al. Change Detection in Remote Sensing Image Data Comparing Algebraic and Machine Learning Methods[J]. Electronic, 2022, 11(3), 431.

[9] Gong M, Cao Y, Wu Q. A neighborhood-based ratio approach for change detection in SAR Images[J]. IEEE Geosci Remote Sens Lett, 2012, 9:307-311.

[10] Hou B, Wei Q, Zheng Y et al. Unsupervised change detection in SAR image based on gauss-log ratio image fusion and compressed projection[J]. Remote Sens, 2014, 7(8):3297-3317

[11] Hachicha S, Chaabane F. Comparison of change detection indicators in SAR images[J]. In: 8th European conference on synthetic aperture radar. EUSAR 2010, 109-112

[12] Bovolo F, Bruzzone L. A theoretical framework for unsupervised change detection based on change vector analysis in the polar domain[J]. IEEE Trans Geosci Remote Sens, 2007, 45:218–236.

[13] Celik T. A Bayesian approach to unsupervised multiscale change detection in synthetic aperture radar images[J]. Signal Process, 2010, 90:1471–1485.

[14] Gong M, Zhou Z, Ma J. Change Detection in Synthetic Aperture Radar Images based on Image Fusion and Fuzzy Clustering[J]. IEEE Transactions on Image Processing, 2012 21(4), 2141-2151.

[15] Zheng Y, Zhang X, Hou B, et al. Using Combined Difference Image and k -Means Clustering for SAR Image Change Detection[J]. IEEE Geoscience and Remote Sensing, March 2014, 11(3), 691-695.

[16] Lin Z K, Liu W, Niu C Y et al. Synthetic Aperture Radar Image Change Detection Based on Difference Image Construction of Log-Hyperbolic Cosine Ratio and Multi-Region Feature Convolution Extreme Learning Machine[J/OL]. Acta Optica Sinica, 2023, 43(12), 1228001.

[17] 王恒涛,张上,贾付文.基于随机森林SAR图像变化检测方法[J].激光杂志,2023,44(06):72-77.

[18] Xuan J Y, Xin Z H, Liao G S. Change Detection Based on Fusion Difference Image and Multi-Scale Morphological Reconstruction for SAR Images[J]. Remote Sensing, 2022, 14(15), 3604

[19] Cui B, Peng Y, Zhang Y H, et.al. Enhanced Edge Information and Prototype Constrained Clustering for SAR Change Detection[J]. IEEE Transactions and Geoscience and Remote Sensing, 2024, 62, 5206116.

[20] Gong M, Su L, Jia M, et al. Fuzzy clustering with a modified MRF energy function for change detection in synthetic aperture radar images[J]. IEEE Transactions on Fuzzy Systems, 2013, 22(1): 98-109.

[21] Moser G, Serpico S, Vernazza G. Unsupervised change detection from multichannel SAR images[J]. IEEE Geoscience and Remote Sensing, 2007, 4:278–282.

[22] Kasetkasem T, Varshney PK. (2002) An image change detection algorithm based on Markov random field models[J]. IEEE Geoscience and Remote Sensing, 2002, 40:1815–1823

[23] Wang F, Wu Y, Zhang Q et.al. Unsupervised change detection on SAR images using triplet Markov field model[J]. IEEE Geoscience and Remote Sensing, 2014, 10:697–701.

[24] Baselice F, Ferraioli G, Pascazio V. Markovian change detection of urban areas using very high resolution complex SAR images[J]. IEEE Geoscience and Remote Sensing, 2014, 11:995–999.

[25] Zhou L, Cao G, Li Y, et.al. (2016) Change detection based on conditional random field with region connection constraints in high-resolution remote sensing images[J]. Remote Sensing, 2016, pp 1–1.

[26] Wang X H, Feng G C, He L J, et.al. Evaluating Urban Building Damage of 2023 Kahramanmaras Turkey Earthquake Sequence Using SAR Change Detection[J]. Sensors, 2023, 23(14), 6342.

[27] Celik T. Unsupervised change detection in satellite images using principal component analysis and k-means clustering[J]. IEEE Geoscience and Remote Sensing,, 2009, 6:772–776.

[28] Liu S, Bruzzone L, Bovolo F et al. Hierarchical unsupervised change detection in multitemporal hyperspectral images[J]. IEEE Geoscience and Remote Sensing, 2015, 53:244–260.

[29] Li H C, Celik T, Longbotham N. Gabor feature based unsupervised change detection of multitemporal SAR images based on two-level clustering[J]. IEEE Geoscience and Remote Sensing, 2015, 12:2458–2462.

[30] Kittler J, Illingworth J. Minimum error thresholding. Pattern Recogn[J]. 1986, 19:41–47

[31] Moser G, Serpico SB. Generalized minimum-error thresholding for unsupervised change detection from SAR amplitude imagery[J]. IEEE Geoscience and Remote Sensing, 2006, 44:2972–2982.

[32] Xue D H, Lei T, Jia X H, et.al. Unsupervised Change Detection Using Multiscale and Multiresolution Gaussian-Mixture-Model Guided by Saliency Enhancement[J]. Earth Observations and Remote Sensing, 2021, 14:1796-1809

[33] Zhang W H, Jiao L C, Liu F, et.al. Adaptive Contourlet Fusion Clustering for SAR Image Change Detection[J]. IEEE Transactions on Image Processing, 2022, 31:2295-2308.

[34] Zhang X Z, Su H, Zhang C, et.al. Robust unsupervised small area change detection from SAR imagery using deep learning[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2021, 173:79-94.

[35] Zhang K Y, Fu X K, Lv X L, et.al. Unsupervised Multitemporal Building Change Detection Framework Based on Cosegmentation Using Time-Series SAR[J]. Remote Sensing, 2021, 13:471

[36] Gong M, Zhao J, Liu J et al. Change detection in synthetic aperture radar images based on deep neural networks[J]. IEEE Trans. Neural Netw. Learn. Syst. 2015, 27, 125–138.

[37] Samadi F, Akbarizadeh G, Kaabi H. Change detection in SAR images using deep belief network: A new training approach based on morphological images[J]. IET Image Process. 2019, 13, 2255–2264.

[38] Li M K, Li M, Zhang P, et al. SAR Image Change Detection Using PCANet Guided by Saliency Detection[J]. IEEE Geoscience and Remote Sensing, 2019, 16(3) 402–406.

[39] 张萌，潘志刚. 基于分层模糊聚类和小波卷积神经网络的SAR图像变化检测算法[J]. 中国科学院大学学报，2023，40(5)，637-646.

[40] Ji L X, Zhao J Q, Zhao Z. A Novel End-to-End Unsupervised Change Detection Method with Self-Adaptive Superpixel Segmentation for SAR Images[J]. Remote Sensing, 2023, 15(7), 1724

[41] Gao Y H, Gao F, Dong J Y, et.al. Synthetic Aperture Radar Image Change Detection via Siamese Adaptive Fusion Network[J]. Remote Sensing, 2021, 14, 10748-10760.

[42] Wang Y. Differentially Deep Subspace Representation for Unsupervised Change Detection of SAR Images[J]. Remote Sensing, 2019, 11.

[43] Gao Y, Gao F, Dong J, et al. Change Detection From Synthetic Aperture Radar Images Based on Channel Weighting-Based Deep Cascade Network[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019, 99:1-13.

[44] Saha S, Bovolo F, Bruzzone L. Building Change Detection in VHR SAR Images via Unsupervised Deep Transcoding[J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 99:1-13.

[45] Zhang X, Su H, Zhang C, et al. Robust unsupervised small area change detection from SAR imagery using deep learning[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2021, 173: 79-94.

[46] Wang R, Wang L, Wei X, et al. Dynamic graph-level neural network for SAR image change detection[J]. IEEE Geoscience and Remote Sensing Letters, 2021, 19: 1-5.

[47] Liu T, Yang L X, Lunga D. Change detection using deep learning approach with object-based image analysis[J]. Remote Sensing of Environment ,2021, 256:112308.

[48] Shu Y J, Li W, Yang M L, et.al. Patch-Based Change Detection Method for SAR Images with Label Updating Strategy[J]. Remote Sens. 2021, 13(7), 1236.

[49] Ma J J, Li D S, Tang X, et.al. Unsupervised SAR Image Change Detection Based on Feature Fusion of Information Transfer[J]. IEEE Geoscience and Remote Sensing, 2023, 20, 4004905.

[50] Ma W, Yang H, Wu Y et al. Change detection based on multi-grained cascade forest and multi-scale fusion for SAR images[J]. Remote Sens. 2019, 11, 142.

[51] Song W Y, Du Y, Wu Y, et al. RF-HoDRF: High-Order Hybrid Discriminative Random Field Improved by Two-Layer Random Forest for SAR Image Change Detection[J]. IEEE Geoscience and Remote Sensing, 2022, 19(10), 4512105.

[52] 王洪菠.极化SAR图像变化检测算法及其快速实现[D].西安电子科技大学，2022.

[53] 张杰.基于样本不平衡学习和结构语义信息的SAR图像变化检测网络[D].西安电子科技大学，2019.

[54] Zhou Z H, Feng J. Deep forest[J]. National Science Review, 2019, 6(01), 74-86.

[55] Breiman L. Random Forests[J]. Machine learning, 2001, 45(1) 5–32.

[56] Geurts P, Ernst D, Wehenkel L. Extremely randomized trees[J]. Machine Learning, 2006, 63, 3–42.

[57] Friedman J H (2000). Greedy function approximation: A gradient boosting machine. Annals of Statistics[J]. ANNALS OF STATISTICS, 2001, 29(5), 1189–1232.

[58] Chen T Q, Guestrin C. Xgboost: A scalable tree boosting system[J]. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, 2016, 785–794.

[59] Ke G, Meng Q, Finley T et al. LightGBM: A highly efficient gradient boosting decision tree[J]. Advances in Neural Information Processing Systems, 2017, 30, 3149–3157.