Using the Keras library to decrypt animal habitats using deep learning methods

The paper considers the use of deep learning algorithms from the Keras library to solve the problem of classifying forest clearings of different ages using remote sensing in the R environment. A rather complicated procedure for installing Keras libraries on a computer is considered in detail. The stages of neural simulation and their variations using the R neuralnet package and the Keras environment are described. Satellite images were decoded in the vicinity of Gomselga village (Karelia) using field survey data. The typical decryption algorithm (classification with learning) was supplemented by a joint multidimensional analysis of the brightness characteristics of the image and field geobotanical descriptions. As a result, 4 sets of reference signatures were formed, corresponding to a particular state of regenerating clearings. The neural network (multilayer perceptron) was configured to recognize these types of plantings, and then performed the classification of the remaining pixels of the image for the entire studied area. Based on the analysis of geobotanical descriptions and satellite data, a grid map was created highlighting four main types of habitats: fresh cuttings, regenerating cuttings, young trees, and deciduous forest. Data processing using Keras algorithms significantly speeds up analysis, and makes it possible to increase the number of layers and neurons and detail the grid. In particular, unlike the algorithms of reference decoding, the proposed approach made it possible to identify the heterogeneity of vegetation within the same-age clearings. The results of the work are used to identify heterogeneous animal habitats and the influence of environmental factors on their spatial distribution and abundance.