Volume XLII-4/W16
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-4/W16, 47–54, 2019
https://doi.org/10.5194/isprs-archives-XLII-4-W16-47-2019
© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-4/W16, 47–54, 2019
https://doi.org/10.5194/isprs-archives-XLII-4-W16-47-2019
© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.

  01 Oct 2019

01 Oct 2019

LOW DATA REQUIREMENTS FRAMEWORK FOR LANDSLIDE SUSCEPTIBILITY PREDICTION USING 3D ALOS PALSAR IMAGES AND NEURAL NETWORKS

S. K. M. Abujayyab and I. R. Karaş S. K. M. Abujayyab and I. R. Karaş
  • Dept. of Computer Engineering, Karabuk University, Demir Celik Campus, 78050 Karabuk, Turkey

Keywords: ALOS PALSAR, Landslide Susceptibility, Neural Networks, Topographic Attributes, Low Requirements Framework

Abstract. Development of landslides susceptibility (LS) predictors based on 3D data is an active area of research in the recent years. Predicting landslides susceptibility maps help to secure human lives and maintaining infrastructures from this risk. Several advanced frameworks proposed with high input data to improve the predictors. The aim of this paper is to develop low data requirement framework for LS predictors development. This framework is only using one input 3D ALOS PALSAR image. The framework has three stages. (A) data pre-processing, (B) deriving explanatory factors, and (C) neural networks training and testing. Exactly. 22 input spatial factors were extracted from ALOS PALSAR image. Extracted factors were utilized to develop the FFNN predictor. The structure of the predictor is 22 factors (input layer) × 150 neurons (hidden layer) × 1 (output layer). Furthermore, 5829 sample points utilized during the training stage, while 745810 points sent to the trained predictor to create LS map. Based on confusion matrix metric, performance accuracy (89.3% training and 82.3 testing), While (95.22% training and 84.7% testing) based on Receiver Operating Characteristic curve. Out of the study area in Karabuk, 3.53 km2 (3.03%) were located in very high susceptibility category. Lastly, the application of the proposed framework showed that it is capable develop low data requirement predictors with high accuracy. Framework provide guideline data for future development in taxing topographic circumstances and large scale of data coverage. In addition, the framework handled the inconsistency in data quality and data updating problem.