Geography An Integrated Approach Pdf

[Valorie Carlee]

Thankyou for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Glacial landforms in mountains are a critical water source for the future, particularly in semi-arid and arid regions7. Studies have revealed that approximately 2.15% of the world's drinkable water is stored as ice in polar and mountainous glaciers, with their residence times ranging from 20 to 100 years8,9. Typically, glacial landforms can be determined using a combination of the visible and shortwave infrared band ratios and Synthetic Aperture Radar (SAR) and topographic datasets10. These techniques, on the other hand, are insufficient to detect and map glacial landforms, which are spectrally indistinguishable from the surrounding paraglacial terrain11. In addition, atmospheric and earth factors may affect the spectral reflectance of glacial landforms12. Therefore, accurate information and inventories of glacial landforms are essential to their management.

Recently, various machine learning-driven methods such as deep learning have been integrated with GEOBIA for the detection and mapping of land use/cover28,29,30,31, gully erosion32,33, and landslide34,35,36. According to the literature review, there is a limited number of research explored the efficiencies of an integrated GEOBIA and convolutional neural network (CNN) to delineate volcanic and glacial landforms.

For volcanic and glacial landforms mapping, we acquired freely available Sentinel-2 imagery (with bands 2 (Blue), 3 (Green), 4 (Red), and 8 (NIR)). Although a high-resolution, freely available DEM for the study area exists (with spatial resolution of 12.5 m), it was necessary to have consistent and comparable datasets that would be applicable to other locations. Due to this, a national topographic map at a scale of 1:25,000 was used to drive DEM. With the spatial analysis carried out in the ArcGIS environment, secondary datasets were generated, including aspect, slope, curvature, and flow accumulation with a spatial resolution of 12.5 m (Fig. 1). In Table 1, we list the characteristics of predisposing variables for volcanic and glacial landforms.

To train the CNN models, we used an inventory map of volcanic and glacial landforms, delineated outlines generated from semi-automated GEOBIA, and ground control points (GCPs). All 935 GCPs (Fig. 7) were collected from the study area with GPS, geomorphological maps, and Google Earth to validate the accuracy of GEOBIA-generated objects and CNN. 70% of these datasets were used to train models, while the remainder (30%) were employed to verify classification accuracy.

An overview of the methodology for detecting and mapping volcanic and glacial landforms is shown in Fig. 2. In the first step, we segmented our datasets using the Multi-Resolution Segmentation (MRS) algorithm in the eCognition software (www. geospatial.trimble.com). Our next step was to generate landform image objects based on geometrical, textural, spectral and contextual features in GEOBIA. To train the landform CNN models, not only generated objects from GEOBIA but also an inventory map of volcanic and glacial landforms as well as GCPs were used. Finally, seven evaluation indexes, including intersection over union (IOU) values, recall (RC), precision (PC), specificity (SP), F-measure (FM), accuracy (ACC), kappa (KP), Fivefold cross validation as well as fuzzy synthetic evaluation (FSE) were employed to validate the accuracy of the classification results.

As shown in Table 4, the segments were selected for volcanic and glacial landforms mapping based on their spectral, textural, and geometrical characteristics. 19 variables were derived from eCognition in order to get as many variables as possible (Table 5). AND fuzzy-based operator then employed in eCognition to classify volcanic and glacial landforms based on fuzzy membership values. Not only GCPs collected by GPS, but also the existing geomorphological map, as well as aerial images, were employed to acquire training data. In sum, 935 sample points were used to identify the most appropriate threshold values for object features and to train CNN models. A rule-based approach is necessary to identify and apply object features to landform classes in GEOBIA. As a result, we incorporated training data along with the efficiency of related spatial and spectral object features for each landform class, obtained through fuzzy threshold values (Table 5). Figure 5 illustrates the performance of some of the training data over selected object-based features.

A convolutional neural network (CNN) is a type of machine learning technique that works with arrays of data, such as one-dimensional signals or sequences as well as two-dimensional visible-light images or audio spectrograms48. Images are consisted of two-dimensional arrays of data, which makes CNN an appropriate tool for image analysis. The CNN is the most common type of deep neural network applied to remote sensing images due to its high generalization capabilities, derived from the features it extracts and its ability to train on extremely large datasets49,50. Neurons are the building blocks of all layers in a neural network. Each neuron represents a convolutional layer aimed at automatic feature extraction from the input image51,52. Figure 6 illustrates the CNN modelling structure for volcanic and glacial landforms.

We used 16 and 9 (Table 5) convolutional layers to train our CNN models for volcanic and glacial landforms, respectively (Table 6). There are several factors involved in each convolutional layer, including a pooling operation, multiple weights, and an activation function. Max-pooling was used with \(2 \times 2\) filters and a two-pixel stride to down-sample the feature maps in the encoder based on a maximum operator by taking the maximum of each \(4 \times 4\) matrix and putting it in the output. Landform detection in this study was done using a \(128 \times 128\) pixel input window. Totally, we applied a twenty-four-layer CNN model for landform detection separately. We fed the twenty-four -layer depth CNN separately with all input window sizes of the training sample patches using nineteen variables (Table 5). In this regard, the input sample patch had \(a \times a \times 19\) units, where \(a \times a\) denotes the size of one layer of sample patches (\(128 \times 128\)), and 19 is all number of the layers required for the analysis. Several convolutions were performed on input using different filters \(\left( 2 \times 2 \right)\) resulting in distinct feature maps. All these feature maps are stacked together to form the convolution layer. By stacking all features along the depth dimension, we generated the final landform outputs volume of \(128 \times 128 \times 19\) by the network by using 24 different filters (one filter per convolutional layer).

Layers of convolution are applied to the valid portions of the image (without any kind of padding) and they are associated with a convolution combined with an activation function to introduce nonlinearity53. There are different activation functions (e.g., sigmoid, softmax, tanh, hyperbolic tangent, ReLU and Leaky ReLU), which are required for the forward propagation and its derivative for backpropagation54. Sigmoid, tanh, hyperbolic tangent and Softmax are typically used in normal neural networks. Rectified Linear Unit (ReLu), on the other hand, is commonly used in CNN algorithms due to their superior performance55. Thus, ReLu function was employed in this study to train landform models. In Eq. (1), the Rectified Linear Unit (ReLU) parameters are defined as follows:

Since training is an iterative process, the loss/cost function is necessary to quantify how good the current state of the network (with specific sets of weights) is. This function is based on the principle of increasing forecast accuracy and reducing errors in the network in order to optimize output at the lowest cost possible56. There are several loss/cost functions for problem-solving in classification, including Mean Squared Error (MSE), Cross-Entropy, and Mean Absolute Error (MAE), and subsequently used Cross-Entropy loss (log loss)57.

This study used ADAM to optimize the results of landform-based models. This function can replace the SGD algorithm and take advantage of AdaGrad and RMSprop, which have better performance in sparse gradients and unstable conditions, respectively61,62. Equations (3) and (4) are defined the ADAM optimizer:

GEOBIA uses fuzzy decision rules and membership values as the basis of object-based classification, which makes it a "soft classifier" approach. Due to the segmentation process, scale regulation, and fuzzy decision rules, it is difficult to assign true or false labels to objects in a binary mode63. As a consequence, we applied Fuzzy Synthetic Evaluation (FSE) for the accuracy assessment of the classification results. Two groups of data, including control point data and the respective rate obtained for each point are used in the FSE to calibrate the overall and per-class accuracy of classified maps using GOBIA according to two steps64. The first step is to compute the classification confidence or magnitude of error for each class using the Difference fuzzy function. To obtain the single accuracy value, the second step weights the Difference function categories. Following these two steps, the degree of confidence in the classification can be calculated based on the ratio of matches between sample and reference data, based on their respective interpretation confidence ratings (ICR), for which default values have been suggested65. A combination of GPS data, control points from Google Earth, high-resolution aerial photographs, and geomorphology maps (scale of 1/25,000) was incorporated in our research as reference datasets.

3a8082e126