Eficient and accurate landslide detection is of great significance for an emergency response to geological disasters. However, detecting landslides from remote sensing images faces two challenges: small objects and class imbalance, and distribution inconsistency. In this paper, the progressive label refinement-based distribution adaptation for the landslide detection framework was proposed. The scale promotion, Lovasz loss, and online hard example mining strategy are adopted to alleviate the class imbalance, and the separated normalization and pseudo label refinement were proposed to encode the statistical inconsistency for reducing the distribution diferences between the training and validation/testing data. The proposed framework has a significant potential for the large-scale global typical natural disaster monitoring rapidly from multi-sensor remote sensing imagery and ranking first place in the validation (F1-score=80.41%) and test leaderboard (F1-score=74.54%) in the LandSlide4Sense competition.
it is laborious, time-consuming, and subjective [5]. Deep learning-based methods make fully automated one-stage Landslide is a worldwide destructive natural phe- landslide extraction possible, and these methods are wellnomenon, usually following an earthquake or heavy rain- reviewed in [5]. However, most of these methods were fall, where thousands of small to medium-sized ground validated in small local regions, and the performance movements occur [1]. Landslides bring serious harm of these models applied directly to a new region in an to society and the economy. Remote sensing technol- emergency is unclear. To promote the development of ogy ofers the possibility of rapid and large-area land the landslide detection field, the LandSlide4Sense comcover monitoring [2, 3], and the detection of globally petition was held and a large landslide dataset, which distributed landslides from multi-source, multi-spectral was collected from diverse geographical regions, is pubremote sensing images using machine learning and com- licly available to help develop a new landslide detection puter vision algorithms facilitates rapid response and algorithm [5]. Landslides detection from large remote management of landslide-generated disasters. sensing imagery will encounter the following two prob
In the early stage of the research, the methods for lems: 1) Small objects and class imbalance and 2) identifying landslides from remote sensing images were Distribution inconsistency. mostly semi-automatic two-stage methods: extracting As shown in Figure.1, small objects and class imbalance discriminative features of landslides through expert are the first challenges of landslide detection. In the real knowledge firstly, and then using SVM or RF for clas- scene, the landslide will have some small branches or sification [ 4]. Although using expert knowledge to con- the landslide itself is relatively small in area, and as a struct discriminative features is transparent and flexible, result, there will be a serious imbalance in the number CDCEO 2022: 2nd Workshop on Complex Data Challenges in Earth of pixels between the landslide and the background, as Observation, July 25, 2022, Vienna, Austria shown in the Figure.1(b) where the number of pixels in * Corresponding author. the background is 49 times that of the landslide in the † These authors contributed equally. training set. Small objects and class imbalance will lead $ 2019206190044@whu.edu.cn (H. Zhao); kingdrone@whu.edu.cn to the problem of lower recall scores. (J. Wang); panyang@whu.edu.cn (Y. Pan); As shown in Figure.2, distribution inconsistency is the (mXa.aWiloanngg0);0z7h@ownghyua.endfeui.@cnw(Ahu..Medau).;cwna(Yn.gxZihnoynug@) whu.edu.cn second challenge of landslide detection. The mean and https://github.com/Hengwei-Zhao96/ (H. Zhao); standard deviation values of the training, validation, and http://junjuewang.top/ (J. Wang) testing sets are counted band-by-band and displayed in 0000-0001-5878-5152 (H. Zhao); 0000-0002-9500-3399 (J. Wang); the Figure.2, where the histogram is the mean and the 0000-0002-6190-4340 (Y. Pan); 0000-0003-3692-6473 (A. Ma); error bars are the standard deviation. Because landslide 0000-0002-0493-3954 (X. Wang); 0000-0001-9446-5850 (Y. Zhong)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License remote sensing images are collected from diverse geoCPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) graphical regions, there are significant inconsistencies in (a) Sentinel-2 imagery with (b) Ratio of the number of pixlandslide els to resize the original images (128 × 128 pixels) into 512 × 512 pixels. Besides, random flip, rotation, and color perturbation are adopted for data augmentation. As we stacked multi-spectral, DEM, and Slope data as inputs, the color perturbation is only applied to spectral data.
As the training and validation (testing) data have inconsistent statistics, the mean values of the pixel values of the data in the source and target domains are significantly diferent. Separated normalization is proposed to reduce the statistical diference between two domains, which takes diferent mean and standard deviation values to normalize the data in the source and target domains, respectively. The mean and standard deviation values were calculated from train and validation/test sets, respectively. Separate normalization is similar to cross-sensor normalization [3], but the domain-specific statistical normalization is performed in the input of the model.
To address the two challenges above, this paper proposed the progressive label refinement framework for domain statistics adaptation, including data preprocessing, model ensemble, model training, model inference, and pseudo label refinement. The overview of the proposed algorithm is shown in Figure. 3.
probability values output by the above three models is taken as the final inference result.
To further align the distributions of the two domains, the progressive label refinement is designed to improve the pseudo-labels. Based on the model prediction, the pseudo labels can be generated from the best models in the ℎ round, using the threshold of 0.7. As for the +1ℎ round, the source samples come from the train set and the target samples are test images with pseudo labels. The pseudo-label generation and domain-adaptation training perform iteratively, progressively refining the test labels.
diverse geographical regions, which consists of training, validation, and test sets containing 3799, 245, and 800 image patches, respectively. Each image patch is a composite of 14 bands that include: multi-spectral data from Sentinel-2 (B1-B12), slope data from ALOS PALSAR, and DEM from ALOS PALSAR. All bands in the competition dataset are resized to the resolution of about 10m per pixel. The image patches have the size of 128× 128 pixels and are labeled pixel-wise. We set batch size as 16 and each model was trained for 20 steps.
The last round refinement results on the validation leaderboard are shown in Table 1. Compared with the baseline, separate normalization significantly improved the accuracy. The selected advanced networks were reifned with several rounds and we ensemble them to obtain the highest F1-sorce=80.41%.
The results from the best models serve as a baseline and achieve F1-sorce=73.07% on the test leaderboard. Similar to the validation development, the label refinement was continuously performed on the test set. The test results in Table. 2 show that the performances of the models are progressively improved as the round increases. Round3 obtains the best result F1-sorce=74.54%.
tribution adaptation for landslide detection framework was proposed. Through multiple rounds of pseudo-label optimization and separately normalization, the performance of the model continues to improve. Our solution ranked first place on the Landslide4Sense challenge. In the future, we will extend the framework into multitemporal images for landslide monitoring.
Intelligence for organizing this competition. This work was supported by National Natural Science Foundation of China under Grant No.42071350, No.42171336 and No.42101327, in part by the Fundamental Research Funds for the Central Universities under Grant 2042021kf0070, and LIESMARS Special Research Funding.