Search published articles


Showing 2 results for Multilayer Perceptron

Ali Jahani,
Volume 6, Issue 2 (9-2019)
Abstract

Risks assessment of forest project implementation in spatial density changes of forest under canopy vegetation using artificial neural network modeling approach
 
Nowadays, environmental risk assessment has been defined as one of the effective in environmental planning and policy making. Considering the position and structure of vegetation on the forest floor, the main role of forest under canopy vegetation cover can be noted in attracting and preventing runoff in the forest floor and reducing subsequent environmental risks. The purpose of this article is forest under canopy vegetation density changes modeling considering forest ecosystem structure and forest management activities as an environmental risk. The main objectives of this study were to: (1) model forest under canopy vegetation density in forest ecosystem to elucidate the ecological and management factors affecting on under canopy vegetation density; (2) prioritize the impacts of model inputs (ecological and management factors) on under canopy vegetation density using model sensitivity analysis and (3) determining the trend model output changes in respond to model variables changes.
In this study, Land Management Units (LMUs) were formed in the region considering ecological characteristics of land. LMUs were mapped out based on Ian McHarg’s overlay technique by ARC GIS 9.3 software. Ecological factor classes of an LMU differ from ecological factor classes of adjacent LMUs (at least in one ecological factor class). The following types of data were solicited for each LMU:
(1) Ecological variables: Altitude or elevation (El), Slope (Sl), Aspect (As), soil depth (SD), Soil Drainage (SDr),Soil Erosion (SE), Precitipation (Pr), Temprature (Te), trees Diameter at Breast Height (DBH), Canopy Cover (CC), and forest Regeneration Cover (RC).
(2) Management variables: Cattle Density (CD), Animal husbandry Dsitance (AD), Road Dsitance (RD), Trail Dsitance (TD), logs Depot Dsitance (DD), Soil Compaction (SC), Torist impacts (To), Skidding impacts (Sk), Logging impacts (Lo), Harvested trees volume (Ha), artificial Regeneration (Re) and Seed Planting (SP).
(3) Forest under canopy vegetation density: The percentage of under canopy vegetation density in each LMU was estimated by systematic random sampling method. In each LMU, a one square meter sample was taken. The average percentage of under canopy vegetation density in sample units of each LMU was calculated and used in the modeling process.
ANN learns by examples and it can combine a large number of variables. In this study, an ANN is considered as a computer program capable of learning from samples, without requiring a prior knowledge of the relationships between parameters. To objectively evaluate the performance of the network, four different statistical indicators were used. These indicators are Mean-Squared Error (MSE), Root Mean-Squared Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R2).
Various MLFNs were designed and trained as one and two layers to find an optimal model prediction for the under canopy vegetation density and variables. Training procedure of the networks was as follows: different hidden layer neurons and arrangements were adapted to select the best production results. Altogether, many configurations with different number of hidden layers (varied between one and two), different number of neurons for each of the hidden layers, and different inter-unit connection mechanisms were designed and tested.
In this research, 129 LMUs were totally selected, then ecological and management variables were recorded in them. In the structure of artificial neural network, ecological and management variables were tagged as inputs of artificial neural network and the percentage of under canopy vegetation density was tagged as output layer. Considering trained networks (the structure of optimum artificial neural network has been summarized in Table1), Multilayer Perceptron network with one hidden layer and 4 neurons in each hidden layer created the best function of topology optimization with higher coefficient of determination of test data (which equals 0.857) and the lowest MSE and MAE (which are 0.866 and 0.736 respectively). Considering the results of sensitivity analysis, ecological and management variables like the forest canopy density, cattle density in forest, soil erosion and soil compaction respectively show the highest impact on forest under canopy vegetation density changes (Fig1).
 
Table1. The structure of optimum artificial neural network in forest under canopy vegetation density

Output Layer First Hidden Layer Network features
Linear Hyperbolic tangent Transmission Layer
Gradient descent Gradient descent Optimization Algorithm
0.7 0.7 Momentum
1 4 Number of Neurons
-0.9 up to 0.9 -0.9 up to 0.9 Normalization
 
Table2. The structure of optimum artificial neural network in test data
MSE MAE RMSE R2 Data The structure of network( the number of neurons)-epoch
0.716 0.678 0.846 0.931 Trainning Tanh(4)-160
0.793 0.703 0.891 0.894 Validation
0.866 0.736 0.931 0.857 Test

 

 
Fig1. The results of sensitivity analysis of artificial neural network model
 
Nowadays, artificial neural network modeling in natural environments has been applied successfully in many researches such as water resources management, forest sciences and environment assessment. The results of research declared that designed neural network shows high capability in forest under canopy vegetation density modeling which is applicable in forest management of studied area. Sensitivity analysis identified the most effective variables which are influencing under canopy vegetation density.
So, to identify hazardous LMUs in study area, we should pay attention to the canopy density of LMUs as the variable with high priority in determination of under canopy vegetation density. We believe that, in hazardous LMUs in forests, we should pay attention to some modifiable factors of LMU, which is cattle density in forest, by timely plan for livestock elimination. The forest under canopy vegetation density assessment model, in forest projects impact assessment, could be a solution in decision making about forest plan structure and implementation of similar projects in similar locations. 
 
Keywords: Forest plan, Environmental impact assessment, Multilayer perceptron, under canopy vegetation, artificial neural network
 


Omid Ashkriz, Fatemeh Falahati, Amir Garakani,
Volume 8, Issue 4 (1-2021)
Abstract

The growth of settlements and the increase of human activities in the floodplains, especially the banks of rivers and flood-prone places, have increased the amount of capital caused by this risk. Therefore, it is very important to determine the extent of the watershed in order to increase risk reduction planning, preparedness and response and reopening of this risk. The present study uses the common pattern of the machine and the classification of Sentinel 2 images to produce land cover maps, in order to construct sandy areas and determine land issues affected by the flood of March 2018 in Aqqla city. Also, in order to check and increase the accuracy of the algorithms, three software indices of vegetation cover (NDVI), water areas (MNDWI) and built-up land (NDBI) were used using images. The different sets of setting of each algorithm were evaluated by cross-validation method in order to determine their effect on the accuracy of the results and prevent the optimistic acquisition of spatial correlation from the training and test samples. The results show that the combination of different indices in order to increase the overall accuracy of the algorithms and to produce land cover maps, the forest algorithm is used with an accuracy of 83.08% due to the use of the collection method of higher accuracy and generalizability than compared to. Other algorithms of support vector machine and neural network with accuracy of 79.11% and 75.44% of attention respectively. After determining the most accurate algorithm, the map of flood zones was produced using the forest algorithm in two classes of irrigated and non-irrigated lands, and the overall accuracy of the algorithm in the most optimal models and by combining vegetation indices (MNDWI) was 93.40%. Then, with overlapping maps of land cover and flood plains, the surface of built-up land, agricultural land and green space covered by flood was 4.2008 and 41.0772 square kilometers, respectively.
 

Page 1 from 1     

© 2024 CC BY-NC 4.0 | Journal of Spatial Analysis Environmental hazarts

Designed & Developed by : Yektaweb