Semi-Automatic Fracture Mapping Using Cellular Neural Networks Applied to ALOS PALSAR 2 Images of the Western Highlands of Cameroon

In Cameroon in general and in the Highlands of Cameroon in particular, there is no fracture map since its realization is not easy. The region’s harsh accessibility and climatic conditions make it difficult to carry out geological prospecting field missions that require large investments. This study proposes a semi-automatic lineament mapping approach to facilitate the elaboration of the fracture map in the West Cameroon Highlands. It uses neural networks in tandem with PCI Geomatica’s LINE algorithm to extract lineaments semi-automatically from an ALOS PALSAR 2 radar image. The cellular neural network algorithm of Lepage et al (2000) is implemented to enhance the pre-processed radar image. Then, the LINE module of Geomatica is applied to the enhanced image for the automatic extraction of lineaments. Finally, a control and a validation of the expert by spatial analysis allows elaborating the fracture map. The results obtained show that neural networks enhance and facilitate the identification of lineaments on the image. The resulting map contains more than 1800 fractures with major directions N20˚ - 30˚, NS, N10˚ - 20˚, N50˚ - 60˚, N70˚ - 80˚, N80˚ - 90˚, N100˚ - 110˚, N110˚ - 120˚ and N130˚ - 140˚ and N140˚ - 150˚. It can be very useful for geological and hydrogeological studies, and especially to inform on the productivity of aquifers in this region of high agro-pastoral and mining interest for Cameroon and the Central African sub-region.


Introduction
Fracture mapping is very important in mining and hydrogeological prospecting.
Indeed, fractures promote the circulation of water and facilitate the location of favorable targets for mineralization exploration. In mining prospecting, a hierarchical mapping of fractures allows to orientate the prospecting during the reconnaissance phase and to bring new ideas on potential traps. In hydrogeology and hydrology, they are the origin of the formation of underground water tables and constitute the zones par excellence of water flow.
In the Western Cameroon Highlands in general, geological maps date from the colonial period [1] [2] [3]. Since then, the identification of fractures from field missions has been limited by the difficulty of access to the region. The main obstacles are generally the large area to cover and other natural factors (rivers, mountains, dense woodlands) that make this approach expensive and time consuming for the state of Cameroon.
Satellite images, with the synoptic view they offer, have proven to be complementary to this difficult work and have made it possible to develop several methods for direct and indirect study of fractures [4]. Indeed, in an image, fractures correspond to rectilinear or curvilinear features that can be perceived on the surface of the Earth's crust and that reflect the presence of deeper phenomena (faults, seams, and geological contacts), generally known by the name of lineaments [5]. The identification of lineaments in satellite images is therefore dependent on the ability of the sensor to detect the slight variations in reflectance associated with these geological phenomena [6].
Several studies in Cameroon [4] [7]- [12] and Central Africa [13] [14] have used satellite imagery to detect lineaments. In the West Cameroon Highlands, a region with high agricultural and agro-pastoral potential and with very rugged terrain and complex geomorphology, only the work of [11] focuses on structural mapping by remote sensing. This work, which is limited to the characterization of the large collapse structures of Ndop, Mapé and Batié, has shown that remote sensing is an irreplaceable source of geomorphology [11]. The lineament extraction approaches frequently used are of two types [15]: 1) the manual approach which is done by photo-interpretation of processed images on one hand; and 2) the automatic extraction approach [16] [17] [18] [19] [20]. The manual technique, which is done by photo-interpretation, is very slow, laborious and often gives subjective results. As for the second approach, which is hardly used in structural studies by remote sensing in Cameroon, it has the disadvantage of presenting the lineaments as small strands. Indeed, during extraction, the lineaments of regional scope are intersected by occlusions [21] due to the relief, the presence of trees and buildings that give them a discontinuous appearance.

Location and Hydrogeomorphological Condition
The study area selected in the Western Highlands of Cameroon (WHPC) has an area of 9000 km 2 ( Figure 1) and is located between 09˚50' and 10˚45' East longi- They therefore provide the majority of the water that drains and infiltrates the region.

Geological and Structural Framework
The geology is very complex [10] [11].

Materials
The Douala East [2] and Douala West [1] sheets Image pre-processing and processing were performed with PCI Geomatica.
The neural networks were trained with MATLAB 2020a. Finally, the digitization, spatial analysis and the realization of the maps were carried out with ArcGIS. International Journal of Geosciences The methodological approach used can be summarized in 4 steps: first, pre-processing operations are performed on the ALOS PALSAT image; the cellular neural network algorithm [21] is trained in MATLAB [28] and applied to improve the pre-processed image; then, the LINE module of Geomatica is applied on the improved image for automatic extraction of lineaments; finally, a control and a manual validation by spatial analysis lead to the realization of the fracturing map.

Image Pre-Processing
The pre-processing consisted in applying on the ALOS PALSAR 2 scenes of the study area, mosaic operations and a Lee filter of size 5 × 5 aiming to reduce speckle and improve the readability of the image while preserving the high frequencies and the geological structures ( Figure 3).

Cellular Neural Network for Lineament Detection
The detection of lineaments on an image consists in finding the local variations of gray levels. One of the simplest methods is to calculate the first derivative of the image called gradient [21]. The lineament corresponds in fact to the local maximum of the gradient, the result of this derivative. The objective of the cellular neural networks used is to calculate the gradient while enhancing the lineaments. In a dynamic of continuity, we have implemented the architecture of the directional cellular neural network with large neighborhood and formed by a large matrix of identical, homogeneous and interconnected cells proposed in previous works [21] [28] [29] and which have demonstrated their relevance to the detection of contours (Figure 4).

Automatic Extraction of Lineaments
To isolate the detected lineaments, the LINE algorithm of Geomatica [30] was

Fracture Mapping and Validation
Firstly, the validation consisted in eliminating by geoprocessing, the linear structures (roads, tracks, plantation limits, high voltage lines…) extracted by the LINE algorithm. Secondly, a multi-criteria interpretation allowed to gather information from digital terrain models, previous works and field missions to make a rigorous selection of fractures that have an appreciable length on the image, represent in several cases the alignment of water drainage and vegetation and coincide with slope breaks or drainage anomalies.     Areas of high and very high density occupy more than half of the study area.

Results
These areas dominate the south and northwest. An area of high density appears in the northeast on the density map by number, while it seems to disappear on the map by cumulative length. In general, the low-density zones are located at the limits of the study area and around the alluvial plains where drainage is intense. Finally, from a geological point of view, the high fracture densities dominate the basement zones and thus allow a better knowledge of the geometry of the aquifers in the study area.

Conclusion
This study has shown that ALOS PALSAR 2 satellite images are well suited to the DOI: 10.4236/ijg.2021.1211056 1066 International Journal of Geosciences study of fractures in the western highlands. In this region where field missions are very expensive, the semi-automatic lineament extraction approach using neural networks can be a palliative to facilitate the realization of a detailed fracture map sketch. Indeed, these networks improve the enhancement of pre-processed images and make the identification and automatic extraction of lineaments easy. The expert's intervention is only necessary for the validation of the extracted lineaments and allows to have a more or less exhaustive fracturing map. However, the number of important parameters to define for the image enhancement and for the extraction is the main difficulty in the application of this method. The algorithms of [21] and [29] work, therefore deserve to be tested in different regions of Cameroon and Africa, experimented extensively in synergy with field studies and integrated into major image processing software to facilitate the realization of geological and structural maps in Africa. The method used in this study has allowed us to identify