A New Mask R-CNN Based Method for Improved Landslide Detection

I. INTRODUCTION

• The authors identify landslide risk to Himalayan regions, where slope failures have caused damage to archaeological sites and loss of life.The introduction names Himachal Pradesh, Jammu & Kashmir, and Uttarakhand among the hazard-prone regions.Source: §I · p. 1
• Deep-learning landslide detection requires large datasets with images from distinct conditions, but dataset collection and annotation are key restrictions.The paper also describes retrieving high-resolution natural-hazard imagery as inconvenient and labor-intensive.Source: §I · p. 2
• Prior work spans remote-sensing monitoring and image-analysis approaches, but the paper focuses on detecting landslides from photographs using a pre-trained CNN.The authors identify ResNet-50 and ResNet-101 as their chosen pre-trained models.Source: §I · p. 2
• The proposed work uses image augmentation and transfer learning to support landslide detection with limited labelled data.It investigates pre-trained models with ResNet-50 and ResNet-101 backbones.Source: §I · p. 2
• The study's objectives include collecting freely available landslide photographs from distinct sources and increasing training volume through augmentation.It also plans fine-tuning with limited photographs and evaluation using both ResNet backbones.Source: §I · p. 2

II. ALGORITHMIC PROCEDURE

• Transfer learning transfers knowledge from a substantial source dataset to a smaller target dataset, allowing the method to use fewer landslide images.The paper identifies the target domain as landslide detection.Source: §II · p. 2, Figure 1 · p. 2
• The model uses COCO-trained weights and trains only the last output layer with labelled landslide images.This setup is used with both ResNet-50 and ResNet-101.Source: §II · p. 3
• Mask R-CNN adds a parallel object-mask prediction branch to Faster R-CNN for pixel-level segmentation, alongside localization and classification.The method uses four ResNet stages in its transfer-learning setup.Source: §II · p. 3

• ResNet-50 and ResNet-101 backbones feed feature maps into anchor-based proposals, RoI Align, and a mask-prediction head for landslide regions.Source: §III · p. 3
• Photographs were gathered from UAV imagery, search engines, and other sources, with images of representative landslide terrain selected by visual inspection.Source: §III · p. 3
• Images were manually assessed, resized to 512×512 pixels, then augmented with Image data generator to expand training data and improve training suitability.Source: §III · p. 3
• VIA annotation defines regions and textual descriptions for five classes: Landslide, Vegetation, Water body, Buildings, and Background.Source: §III · p. 4
• Two dataset splits assess consistency: Set A has 101 training, 28 validation, and 31 test images; Set B has 62, 28, and 31, respectively.Source: §III · p. 4
• The model uses MS COCO pretrained weights for ResNet-50 and ResNet-101, a 0.8 objectness threshold, and a 0.001 learning rate.Source: §III · p. 4

IV. SIMULATION RESULTS

• Loss decreased as landslide detection accuracy increased for both ResNets, but overtraining increased loss and reduced accuracy.Table I is identified as reporting overall loss values for ResNet-50 and ResNet-101.Source: §IV · pp. 4–5, Table I · p. 4
• Detected objects receive rectangular bounding boxes when detection accuracy exceeds 80%, and the text says accuracy generally ranges from 95% to 100%.Precision, Recall, and F1 are used for assessment; the passage defines TP, FP, and FN for those measures.Source: §IV · p. 5
• The authors report that instance segmentation is more beneficial than semantic segmentation for distinguishing samples with the same classes.Mask R-CNN generates colored masks for different objects; Faster R-CNN is described as requiring less inference time.Source: §IV · pp. 5–6, Figure 7 · p. 6, Figure 8 · p. 6
• For ResNet-101, Dataset A yielded 1.00 correctly predicted positive observation, 0.93 recall, and 0.97 classification accuracy.The results are reported alongside Table II for Dataset A.Source: Table II · p. 5, §IV · p. 6
• ResNet-50 correctly predicted positive observations at 1.00 on both datasets, with recall of 0.87 on Dataset A and 0.73 on Dataset B.Its classification accuracy was 0.93 on Dataset A and 0.85 on Dataset B.Source: §IV · p. 6
• ResNet-101 recall was 0.93 on Dataset A and 0.87 on Dataset B, while its classification accuracy was 0.90 on Dataset B.The passage reports classification accuracy of 0.97 for ResNet-101 on Dataset A.Source: §IV · p. 6

V. DISCUSSION

• Obtaining sufficient high-resolution training samples remains challenging for deep-learning landslide detection.The authors describe building such models as significantly complex.Source: §V · p. 6
• The approach uses a pre-trained Mask R-CNN, COCO weights, and landslide-labelled images to train the output layer, with ResNet-50 and ResNet-101 backbones.The work uses image augmentation and two datasets; the reported best result was 31 detections with ResNet-101 using 160 images, with 1.00 Precision, 0.93 Recall, and 0.97 F1 score.Source: §V · p. 6
• The authors report that their hybrid Mask R-CNN outperformed the reviewed methods, with comparisons spanning landslide studies in different regions.They contrast their detection and classification assessment with a prior deep-learning and transfer-learning study that used only a confusion matrix for acquisition.Source: §V · pp. 6–7, Table VI · p. 7

VI. CONCLUSION

• ResNet-101 performed better than ResNet-50, with 1.00 Precision, 0.93 Recall, and 0.97 F1 measures on Dataset A, where 31 landslides were detected.The method combines a pre-trained Mask R-CNN with transfer learning.Source: §VI · p. 7
• Manual annotation of training images was the main problem identified by the authors, who describe it as very time-consuming.They say future work will attempt object classification instead of detection to improve accuracy and address manual annotation.Source: §VI · p. 7
• Future work includes object classification for multiple terrain types and improving the traversal scheme to monitor loss values.The authors state that the classification approach would not be based on masked instances.Source: §VI · p. 7