Deforestation Prediction Using Neural Networks and Satellite Imagery in a Spatial Information System
DDeforestation Prediction Using Neural Networks andSatellite Imagery in a Spatial Information System
Vahid Ahmadi
Tarbiat Modares University
Abstract
Deforestation, as one of the challenging environmental problems in the world,has been recorded the most serious threat to environmental diversity and oneof the main components of land-use change. In this paper, we investigate spa-tial distribution of deforestation using artificial neural networks and satelliteimagery. Modeling deforestation can be conducted considering various fac-tors in determining the relationship between deforestation and environmentaland socioeconomic factors. Therefore, in order to ascertain this relationship,the proximity to roads and habitats, fragmentation of the forest, height fromsea level, slope, and soil type. In this research, we modeled land cover changes(forests) to predict deforestation using an artificial neural network due to itssignificant potential for the development of nonlinear complex models. Theprocedure involves image registration and error correction, image classifi-cation, preparing deforestation maps, determining layers, and designing amulti-layer neural network to predict deforestation. The satellite images forthis study are of a region in Hong Kong which are captured from 2012 to2016. The results of the study demonstrate that neural networks approachfor predicting deforestation can be utilized and its outcomes show the areasthat destroyed during the research period.
Keywords:
Land-cover change, deforestation, Neural Network, GIS
1. Introduction
Deforestation is an important issue that has received considerable atten-tions in many different disciplines. As shown in previous studies cite, this
Email address: [email protected] (Vahid Ahmadi) a r X i v : . [ c s . C Y ] D ec henomenon has a negative impact on regional hydrology, large-scale andlong-term climatic systems, global biogeochemical cycles, and extinction ofanimal species [1], [2], [3] [4] [5]. Despite its importance, most countries,including Iran, do not have detailed statistics on the extent of deforestation.In Mexico, authors have reported deforestation rates of 0.3-0.5% per year[6],[7],[8],[9],[10],[11]. Also, one study estimated the rate of deforestation tobe about 0.3% and 0.8% per year for the temperate and tropical areas. Thisstudy showed that 84,000 km2 of forest cover were destroyed between 1976and 2000 [8], [11].Geospatial Information System(GIS) and Remote Sensing are two essen-tial technologies which can be used to implement and couple spatial analysesand machine learning algorithms for predicting deforestation. Spatial analy-ses cover a wide spectrum of different spatial problems in various scales. Itcan be used for very small scale movement pattern analysis, indoor navigationand obstacle detection [13], [14],hydrology, way-finding and path planning[15], large-scale accessibility measurement [16], [17],[18], urban informatic,public health, environmental preservation, agriculture, military, and so on.In this research we utilize GIS by integrating it RS and implementing neuralnetwork in it.Although the main driving factors in deforestation are known, it is diffi-cult to assess their contribution to deforestation and to our best of knowledge,there is no clear understanding of the interaction of these factors. Simulationof land cover and land use change plays a crucial role in management naturalresources as well as in academic research. In deforestation, the developmentof models is accomplished by several useful factors [19], [20], [21]: Providinga better understanding of how deforestation factors play a role Production ofa future scenario for deforestation Forecast of forest destruction In order tosupport the design of responsible forestry policy The purpose of this studyis to develop a simple spatial model that can predict deforestation usingartificial neural networks [22],[23],[24].
2. Combining the potential of deforestation in a database of spatialinformation systems
Various factors in determining the relationship between deforestation andenvironment and socioeconomic factors can be considered. Effective factorssuch as distance from roads, distance from habitats, altitude, slope, soil2ender and patterns of forest fragmentation can be mentioned. Several im-portant factors that describe deforestation include:1. Height2. Slope3. Soil type4. The shortest distance from the nearest one5. The shortest distance from the nearest habitat6. The shortest distance from the nearest edge of the forest or non-forest7. The fragmentation of the forest covering the surrounding and immedi-ate area of the siteForest fragmentation can be estimated using two indicators • Forest cover index • Matron indexThe forest cover index, expressed as percentages, shows the calculatedpixel count of the forest in 3 ×
3, 9 × ×
15 pixels windows. In thisstudy, a 3 × × M = N F − NF √ N F √ N (1)Where N F is the number of boundary pixels between forest and non-forest,NF is the number of pixels in the forest and N is the total number of pixels[25].To determine the number of suitable variables for entering the artificialneural network, the correlation coefficients between the variables can be cal-culated [26],[27]. Variables such as altitude, forest cover and distance andproximity to residential areas were selected as network inputs.
3. Multilayer Perceptron Neural Network
Selected variables were used to describe the network’s training data. Inthe model, the data were divided into two categories: training set and testset. The training algorithm does not use under any circumstances the testdata to adjust network weights because the test set is used to detect network3xecution errors and stop network training if over-learning occurs, so thereshould be no dependence on network parameters.We used a neural network with similar configuration to [ ? ] in which thetraining process was tested with Levenberg-Marquadt and Back-propagationalgorithms. Deciding on the number of hidden layers experimentally by test-ing various choices that provide the best balance between the bias and thevariance [28], [29], [30]. Finally, a network with three levels of input layer,hidden layer and output layer was designed ( Figures 1 ). Figure 1: Schematic diagram of the Network
Using layers in GIS environment, the network was trained with existingdata and then used to prepare a deforestation risk map. The network outputis a quantity that expresses the natural tendency of each pixel for deforesta-tion. So, the result is a fuzzy deforestation map that shows the probabilityof deforestation, which ultimately is classified.
4. The study area and the stages of the work
The studied area is part of the forested area in Hong Kong, China, whichis located in the southeast of Asia between N ◦ (cid:48) – N ◦ (cid:48) and E ◦ (cid:48) – E ◦ (cid:48) ( Figures 2 ). 4 igure 2: Satellite image of the study area
Two SPOT images of the same area, taken on 14/01/2012 and 20/12/2016,were used along with ground control points for geometric correction and re-gression, and a 90-meter DEM of the region, which was provided by theSRTM database. Stages of work in the model The study consists of six stepsas follows:1. Registry and image correction2. Classification of images into three classes of forest, wetland and urbanareas3. Preparation of deforestation maps obtained from overlapping forestcover maps multiple times4. Obtaining appropriate layers and their configuration for modeling5. Modeling (Neural Network Training)6. Simulation (providing a natural tendency map for deforestation thatpredicts deforestation for the next period)To carry out the steps, first, the images taken in the ArcGIS environmentwere retrieved using Extention, Georefrencing, as shown in
Figures 3 (a)and 3 (b) . 5 a) 2012 (b) 2016
Figure 3: Georeferenced SPOT satellite images of the area in two snapshots
The images were then embedded in the ERDAS software and the imageswere classified using the algorithm most similarly to forest, sea, and urbanclasses and loaded again into the ArcGIS environment. The clustered imagesare shown in
Figures 4(a) and 4(b) (a) 2012 (b) 2016
Figure 4: Classified images into forest, sea, and urban areas
From Fig. 4 (a) and Fig. 4 (b), a layer is derived for forest areas. In the6 a) 2012 (b) 2016
Figure 5: Reclassification into forest and non-forest classes output images, only two values are zero (forests) and one (non-forest). Theoutput of this process for the two years studied is shown in Figures 5 (a) and5 (b):These layers were used to generate the forest cover index. For this pur-pose, these layers were outputted in ASCII format and MATLAB softwarewas written for this purpose, and its output was returned to the ArcGISenvironment again. This output resulted in two images depicted in Figures6 (a) and 6 (b).The two layers of the forest cover index are the first line of entry for thenetwork. Now the production of the second layer, distance and proximityto the cities, has gone as the second line of input of the network. For thispurpose, a layer is first produced which consists of two classes of urban andnon-urban, the image of these two classes in Fig. 7 (a) and Fig. 7 (b).Now with the help of the Spatial Analyst functions in ArcGIS software,we produce the layer for entering the network, distance from urban areas. InFigure 8 (a) and Figure 8 (b), the image of this layer is well visible.Finally, the last entry as the third line of the vector input is the height.As mentioned earlier, the region DEM (Fig. 9) was used after a series ofprocesses to extract a common area. This DEM contains information innon-marine areas with elevations ranging from 943 to 80 m.7 a) 2012 (b) 2016
Figure 6: Generated forest cover index for 2012 and 2016 (a) 2012 (b) 2016
Figure 7: Reclassification into urban and non-urban classes a) 2012 (b) 2016 Figure 8: Distance layer - computed distance from urban areasFigure 9: Schematic diagram of the Network igure 10: Schematic diagram of the Network Once the network input data is ready, the training data must be gener-ated. For this purpose, there has to be a process to explore deforestationduring these years. In this process, the areas covered by forest cover in 2012and in 2016 will be covered outside of the forest, amounting to one, andif they maintain their forest cover, zero is allocated to non-refugee areas in2012, NoData will be allocated. The output of this process is shown in Figure-10. By taking the ASCII output from this map, which indicates actual de-forestation, part of it, along with the corresponding entries for 2012, is givenas training data for network education. The other part of the data that wasnot involved in the training process was used to test the network, resulting ina precision of 98Network convergence was performed by setting the networkparameters as follows to achieve the desired accuracy. (Figure-11)10 igure 11: Schematic diagram of the NetworkFigure 12: Schematic diagram of the Network
Once the network parameters are set, it is simulated. To do this, we haveto retrieve the data from 2016 to the network to get results about the future.The result of this stage is a de industrial risk map divided into three classes,class one related to areas without deforestation, the second class relates toareas relatively prone to deforestation and to some extent sustainable, and11he third class relates to areas at high risk of deforestation and Extremelyunstable. In Figure 12, a final map of the prediction of deforestation in thestudied area is presented.
5. Conclusion
In this paper, the ability of artificial neural networks to model land coverchanges as a powerful method for investigating deforestation was investi-gated. However, the development of models for deforestation and land usechange due to the high dependence of models on large environmental factors,due to changing climatic and economic conditions, is practically impossibleto accurately predict. Therefore, in this research, using the neural networksof a part of the forest that was destroyed was accurately predicted. Thismodel can be used by environmentalists and managers to develop targetedpolicies to control the ecological and social impacts of deforestation.
References [1] W. Steffen, K. Richardson, J. Rockstr¨om, S. E. Cornell, I. Fetzer,E. M. Bennett, R. Biggs, S. R. Carpenter, W. de Vries, C. A. de Wit,et al., Planetary boundaries: Guiding human development on a changingplanet, Science 347 (6223) (2015) 1259855.[2] C. D. Allen, D. D. Breshears, N. G. McDowell, On underestimationof global vulnerability to tree mortality and forest die-off from hotterdrought in the anthropocene, Ecosphere 6 (8) (2015) 1–55.[3] B. L. Turner, D. Skole, S. Sanderson, G. Fischer, L. Fresco, R. Leemans,Land-use and land-cover change: science/research plan, in: UnknownHost Publication Title, International Geosphere-Biosphere Programme,Stockholm; Report, 35, 1995, pp. –.[4] C. Howarth, D. Viner, S. Dessai, C. Rapley, A. Jones, Enhancing thecontribution and role of practitioner knowledge in the intergovernmentalpanel on climate change (ipcc) working group (wg) ii process: Insightsfrom uk workshops, Climate Services 5 (2017) 3–10.[5] T. W. Crowther, D. S. Maynard, J. W. Leff, E. E. Oldfield, R. L. Mc-Culley, N. Fierer, M. A. Bradford, Predicting the responsiveness of soil12iodiversity to deforestation: a cross-biome study, Global change biology20 (9) (2014) 2983–2994.[6] P. D. Richards, R. T. Walker, E. Y. Arima, Spatially complex landchange: The indirect effect of brazil’s agricultural sector on land use inamazonia, Global Environmental Change 29 (2014) 1–9.[7] T. Newbold, L. N. Hudson, S. L. Hill, S. Contu, I. Lysenko, R. A. Senior,L. B¨orger, D. J. Bennett, A. Choimes, B. Collen, et al., Global effectsof land use on local terrestrial biodiversity, Nature 520 (7545) (2015)45–50.[8] K. S. Kibret, C. Marohn, G. Cadisch, Assessment of land use and landcover change in south central ethiopia during four decades based onintegrated analysis of multi-temporal images and geospatial vector data,Remote Sensing Applications: Society and Environment 3 (2016) 1–19.[9] J. M. Faggin, A. Offermans, Sustainable forest management as a po-tential integrative approach in international public policy, SustainableDevelopment Research at ICIS (2016) 189.[10] B. A. Margono, P. V. Potapov, S. Turubanova, F. Stolle, M. C. Hansen,Primary forest cover loss in indonesia over 2000-2012, Nature ClimateChange 4 (8) (2014) 730–735.[11] C. Pelletier, S. Valero, J. Inglada, N. Champion, C. Marais Sicre,G. Dedieu, Effect of training class label noise on classification perfor-mances for land cover mapping with satellite image time series, RemoteSensing 9 (2) (2017) 173.[12] P. Gharani, B. Suffoletto, T. Chung, H. A. Karimi, An artificial neuralnetwork for movement pattern analysis to estimate blood alcohol contentlevel, Sensors 17 (12) (2017) 2897.[13] S. Nissimov, J. Goldberger, V. Alchanatis, Obstacle detection in a green-house environment using the kinect sensor, Computers and Electronicsin Agriculture 113 (2015) 104–115.[14] P. Gharani, H. A. Karimi, Context-aware obstacle detection for nav-igation by visually impaired, Image and Vision Computing 64 (2017)103–115. 1315] G. Costante, C. Forster, J. Delmerico, P. Valigi, D. Scaramuzza,Perception-aware path planning, arXiv preprint arXiv:1605.04151.[16] S. N. Zenk, A. J. Schulz, B. A. Israel, S. A. James, S. Bao, M. L. Wilson,Neighborhood racial composition, neighborhood poverty, and the spatialaccessibility of supermarkets in metropolitan detroit, American journalof public health 95 (4) (2005) 660–667.[17] W. Luo, Y. Qi, An enhanced two-step floating catchment area (e2sfca)method for measuring spatial accessibility to primary care physicians,Health & place 15 (4) (2009) 1100–1107.[18] P. Gharani, K. Stewart, G. Ryan, An enhanced approach for modelingspatial accessibility for in vitro fertilization services in the rural Mid-western United States, Applied Geography 64. doi:10.1016/j.apgeog.2015.08.005doi:10.1016/j.apgeog.2015.08.005