APPLICATION OF ARTIFICIAL NEURAL NETWORK TECHNOLOGY IN WATER COLOR REMOTE SENSING INVERSION OF INLAND WATER BODY USING TM DATA

Size: px

Start display at page:

Download "APPLICATION OF ARTIFICIAL NEURAL NETWORK TECHNOLOGY IN WATER COLOR REMOTE SENSING INVERSION OF INLAND WATER BODY USING TM DATA"

Wilfred Bryant
5 years ago
Views:

1 APPLICATION OF ARTIFICIAL NEURAL NETWORK TECHNOLOGY IN WATER COLOR REMOTE SENSING INVERSION OF INLAND WATER BODY USING TM DATA J. P. Wang *, S. T. Cheng, H. F. Jia The Department o Environmental Science and Engineering, Tsinghua University, 00084, Beijing, China Wjp00@mails.tsinghua.edu.cn KEY WORDS: Artiicial Neural Network Model, Water Color Remote Sensing, Extraction, Lake, Landsat Image ABSTRACT: For water color remote sensing study o inland water body, terrestrial satellite is oten a good data source, because it has high spatial resolution. However, the precision o water color remote sensing inversion limits its application to water environmental monitoring and pollution analysis. This paper irstly studied traditional regression arithmetic and ound that it was diicult to extract a good combination to construct the regression model. In order to acquire good inversion results, the paper introduced an advanced nonlinear science, artiicial neural network technology. On the basis o satellite synchronous monitoring experiment, a BP neural network model was constructed to inverse SS, COD Mn, DO, T-N, T-P and chl-a rom Landsat TM data. The accuracy was acceptable and the relative error could be controlled below 5%. Moreover, the reasons o simulating error, ways o improving model and applications o the model were also analyzed in details. The result o this research showed that based on a small-scale o satellite synchronous experiment, the model could be applied successully in investigation, analysis and estimation o water quality. It comes through long time that water quality data collection oten depends on traditional monitoring, which needs much time and labor. So it is impossible to realize real-time and quick data acquiring. Along with continuous development o environmental inormation technology, water color remote sensing is applied more and more widely in the water quality monitoring o oceanic, coastal and inland water body because it has many advantages, such as wide range, synchronization and low cost o data collecting (Campbell, 988; Claudia, 00; Zhao, 000). However, in order to utilize water color remote sensing technology more deeply and widely, there are many aspects to be improved: water color remote sensor technology, atmospheric correction and inversing model, which are in accordance with many scientiic ields, and this paper paid more attention to inversing arithmetic study.. OVERVIEW Researches o water color remote sensing began at 90s, which had given correct explanation o sea color and begun to study optical ield o water body (Shuleikin, 933). But only ater spatial technology occurred, water color remote sensing developed truly and quickly. Morel & Prieur classiied water body as two types: Case I water body, which is Open Ocean; Case II water body, which is coastal, estuary and inland water body (Morel, 977). Now the inversion research o Case I water body is correspondingly mature and the accuracy o inversing model is relatively good, since which component is mainly chlorophyll and has little suspended solids (SS). And that o Case II water body is very diicult due to the interaction o many water components, such as SS, chlorophyll and yellow substance. Inversion research o Case II water body is a hot issue currently. Aiming at Case II water body inversion model, both theoretical model (Bricaud, 986; Cheng, 00) and empirical model (Chen, 996; Ekstrand, 99; Kuang, 00) is getting along in recent years. And TM or ETM data was used by most researches (Chen, 996; Cheng, 00; Zhan, 000). At present, the theoretical research still cannot be applied to practical inversion, but it can provide useul inormation to direct and modiy empirical inversion arithmetic. Empirical model is oten the main choice o quantitative calculation. How to improve its precision has become a key issue. Water color remote sensing satellites, such as SeaWiFS and OCTS, have high spectral resolution and good optical attributes o water body, which are ideal data sources or water color remote sensing research, but in practice, the small spatial resolution oten limits their application in inland water body. Terrestrial remote sensing satellites, such as landsat7, which has good spatial resolution (about 30meters), is oten adopted as data sources, but spectral resolution o terrestrial satellite is a little low and can t relect optical attributes o water body well. All these reasons make it more diicult to identiy suspended solids, chlorophyll and yellow substances, and also limit the application o water color remote sensing in inland water bodies. Artiicial neural network (ANN) technology is a kind o nonlinear science developed rom 980 s, which tries to simulate some basic attributes o people, such as sel-adapting, sel-organizing and ault tolerance. ANN has been used in many ields, such as mode identiication and system simulation. Integrating water color remote sensing and characteristics o ANN, the paper hoped that artiicial neural network model could perorm the research o water color remote sensing inversion well. For all these, the paper drew the ollowing research plan (Figure ). * Corresponding author.

Choose study area Synchronous monitoring Image data pretreatment radiometric correction Geometric correction atmospheric correction whose precision was better than one pixel.

2 Choose study area Synchronous monitoring Image data pretreatment radiometric correction Geometric correction atmospheric correction whose precision was better than one pixel. The ollowing part discussed the procedure o atmospheric correction... Atmospheric Correction PCI, a commercial image processing sotware package, provides a set o atmospheric correction tools or sensors o TM, MSS and SPOT, such as ATCOR0, ATCOR and ATCOR. The low chart o atmospheric correction is showed as Figure 3. Construct inversing model Analysis o results Empirical regression model Artiicial neural network model Original Data: TM/MSS or SPOT3/4 ATCOR0 ATCOR Deine aerosol optical depth: VISIBILITY Create relectance image: Using ATCOR0 to give VISIBILITY Figure. The Frame o Research Plan Relectance Image Without Adjacency Eect. DATA COLLECTION AND PROCESSING FAV Lowpass Relectance Image Lowpass ilter Create lowpass relectance image. Synchronous Monitoring Water quality monitoring experiment synchronized with Landsat satellite was perormed in PoYang Lake in July 8, 00. The locations o sampling points were set as Figure. ATCOR Relectance Image With Adjacency Eect Create improved relectance image with adjacency eect Figure 3. Flow chart o atmospheric correction provided by PCI JiaoTang LiaoNan NanKuang Sampling Points PoYang Lake DuChang The theory o this method is given by Richter (Richter, 990; Richter, 996), whose main idea is that according to standard atmospheric categories, atmospheric dispersion has been calculated in dierent aerosol types, dierent sun zenith angles, dierent altitudes and dierent atmospheric visibilities, the results o which are stored in a directory, like as look-up table. In actual application atmospheric correction is perormed according to this table. Categorizing basis o Richter method comes rom middle resolution atmospheric transmission model - MODTRAN. Its arithmetic also considers and corrects the adjacent eect o ground relection. Figure. Distribution o sampling points There were ten sampling points, at which SS, chl-a, TN, TP, COD Mn, DO, temperature and ph were monitored. DO, temperature and ph were measured in situ, and the others in laboratory. Data o longitude and latitude were obtained by GPS at each point in situ.. Remote Sensing Data Remote sensing data adopted synchronous Landsat 7 ETM+ data or its good spatial resolution, which path/row numbers were /40. The satellite image was clear and cloud-ree... Pre-processing o Remote Sensing Data The remote sensing data needs several steps o pre-processing beore an inversing model is applied, which include radiometric correction, geometric correction and atmospheric correction. Commonly the purchased image has been processed by radiometric correction and original geometric correction, so jobs that users need to do are accurate geometric correction and atmospheric correction. The accurate geometric correction in this study was accomplished by ground control points (GCPs), 3. EMPIRICAL REGRESSION MODEL Empirical Regression Model oten sets remote sensing data (atmospheric correction or not) as independent variables and concentration o water quality components as dependent variable(s) to construct their relative equations. In order to review eects o empirical regression models, this study designed ollowing combinations, and detailed description was showed in the previous study (Kuang, 00). Atmospheric correction: yes or no. two cases. Independent variables: 89 kinds o remote sensing band combination, such as R, R+R3, (R3*R5)/ln(R*R) and so on (Kuang, 00). Dependent variables: SS and chl-a. 3. Results o Regression Model The datasheet o water quality monitoring and remote sensing data were showed in Table. Calculating the relative coeicient o every above combination (the total mounts o combination were 356), results that had good correlation were listed in Table.

3 Atmospheric correction No Yes 3. Analysis Table showed that: dependent variables good band combinations relative coeicien t SS R/ln(R3) 0.9 ln(r+r3)/ln(r) 0.5 Chl-a ln(r*r3)/ln(r) 0.57 R/ln(R3) 0.4 SS (R+ln(R))/ln(R3) 0.48 Chl-a ln(r+r3)/ln(r) 0.64 Table. Results o Regression analysis Most combinations had bad relative coeicients between concentration o water quality component and remote sensing data. As Table, only six combinations were extracted rom the total 356, but their relative coeicient were still not good. Relative coeicient could be improved a little ater atmospheric correction. It was hard to build an applicable regression model or quantitative analysis o Case II water body. The main reason was: Water components interacted each other strongly. Relations between image bands and water components were crossed and non-corresponding relative. The wider band width also decreased their relativities. 4. THE ARTIFICIAL NEURAL NETWORK INVERSING MODEL Theoretical research has proved that i an ANN includes biases, at least a S-style cryptic layer and a linear output layer, it can approach any rational unction (Cong, 998). And theoretical research o water color remote sensing indicates that remote sensing data are correlated with water body components and their concentrations directly ater atmospheric eect is removed *. However, because water body ingredients aect each other, traditional inversing method, such as building a relative unction between band data and components cannot solve it. The artiicial neural network is suitable to simulate such complicated relationship. Thereore, theoretically, it is easible to use ANN in water color remote sensing inversing research. 4. Model Structure Figure 4 showed a concept chart o artiicial neural network structure used in this study. Every input node represented a TM band, and these input data were distributed to each node o cryptic layer to operate. Output values o cryptic layer were inputted into output layer and operated again. The output values o output layer were parameters interested by users. * Kuang, C., 999. Identiication o Water Color Remote Sensing Theoretical Model and Several Aspects in Its Solution. Tsinghua University doctoral dissertations, pp. 5~86. Input First layer Second layer iw, iw, lw, lw, a = (IW * p + b ) a = (LW, *a + b ), Figure 4. The structure o artiicial neural network In detail, the paper adopted a two layer BP artiicial neural network. p i (i=r), r=5, represented ive inputs which were band to band5 o TM image data; iw, represented the weight between inputs and neural cells o irst layer. lw, represented the weight between outputs o irst layer and neural cells o second layer. b and b were biases o neural cells o irst and second layer respectively. was transer unctions. oten adopted S-style transer unction, in this paper the model used hyperbolic tangent S-style transer unction. adopted linear transer unction. Theoretical research has proved that an ANN, which has biases and at least a S-style cryptic layer and a linear output layer, can approach any rational unction (Cong, 998); a and a were output values o irst and second layer (see also Figure 4): a a = = (IW, (LW, * p + b *a ) + b m, the number o neural cells in the second layer, was determined by amounts o object o ANN model object. The paper hoped that this ANN model could simulate SS, COD Mn, DO, TP, TN and chl-a, so m=6; s, the number o neural cells in the irst layer, was determined by neural network training process. According to results o the paper, s was equal to 0. ) 4. Training Results and Its Analysis 4.. Data o Model Training and Veriying Generally the input data should be normalized in order to improve training eiciency and accuracy o neural network (Zhan,000). Table gave the input data and object values.

4 Input Object Item Training samples Veriying samples band band band band band SS CODmn DO T-P T-N chl-a/ug.l Table. Original data o the artiicial neural network model/mg.l - In the training process, size o s, which was amounts o neural cell in cryptic layer, was justiied gradually. Ultimately s was conirmed as 0 by analyzing accuracy o veriying samples, training time and iterative times. The veriying results were showed in Table 3 ater testiying samples were inputted to the having been trained ANN model. Water quality parameters Inversing results Synchronous monitoring results Inversing errors SS COD Mn DO T-P T-N chl-a ug.l Table3. The results o inversion and validation o the model/mg.l - It could be concluded that inversing errors could be controlled below 5%, except or SS and chl-a in second veriying data, which exceeded 30%. The results were satisactory. 4.. Result Analysis Results indicated that the artiicial neural network was well able to inverse water quality parameters rom remote sensing image: () Inversing eect o the model was good (Table 3). Inversing errors could be controlled less than 5%, except or SS and chla in second veriying data. I some measures as ollows were taken in data collecting, processing and analyzing, inversing precision could be improved urther. () Ability o the model in simulating complicated relations was strong. ANN could realize nonlinear mapping between input and output parameter dimensions by the way o adjusting weights and biases o neural nodes (Cong, 998). Results indicated that the ANN model had good simulating eects. (3) The model inversed multi-water quality parameters simultaneously. The paper gave a model that SS, CODMn, DO, TP, TN and chl-a could be inversed rom TM image using a trained ANN in the same time. Reasons o one or two parameters having higher errors might be: synchronous experiment s internal errors, or example, boat stirring in shallow water or laboratory error; little sampling points were placed, which might result in water quality parameters bad-proportioned distribution in their concentration ranges; processing errors o remote sensing data, especially atmospheric correction errors; structure errors o the model. The research showed that the irst item error could be avoided or decreased by some steps, such as sampling ater a moment o boat stopping, or increasing number o samples or each point; the second item error could be removed through steps o increasing sampling points or making these points reasonable arrangement; the third item error was diicult to remove. I it was possible, some optical experiments should be perormed at the same time o satellite synchronous monitoring in order to decrease atmospheric correction errors. The last item error was inevitable, but adjusting neural network structure or comparing various amounts o neural cells could reduce the error. 5. CONCLUSIONS Literature review indicated that water color remote sensing inversion oten adopted empirical model or limits o water color remote sensor technology and atmospheric correction arithmetic. This paper analyzed traditional regression model through combining dierent bands and operations. The relative coeicients between remote sensing data and water quality data was not good and could not satisy the requirements o application. Analysis indicated that it was resulted rom the interaction o many water quality components, such as SS, chlorophyll and yellow substances.

5 Both theoretical study and experimental analysis indicated that the ANN technology was easible to water color remote sensing research, and the model had strong ability to simulate complicated inversing relation o second type water body. On the basis o satellite synchronous monitoring experiment, a BP neural network model was constructed, by which concentrations o SS, CODMn, DO, T-N, T-P and chl-a were inversed rom Landsat TM data and the accuracy was acceptable, the relative error could be controlled below 5%. And reasons o simulating error, ways o improving model and applications o the model were analyzed in details. The result o this research showed that based on a small-scale o satellite synchronous experiment, the model could be applied successully in investigation, analysis and estimation o lake water quality. Shuleikin, V. V., 933. Data on the optics o a strongly scattering medium, applied to sea water, og and cloud. Geoizika, 3, pp Zhao, B. Y., He, B., Zhu, Y. Y., etc., 000. Remote sensing water quality model o suspended sediments in the dianchi lake water bodies. Environmental Science and Technology, 3(94), pp. 6-8(in Chinese). Zhan, H. G., Shi, P., Chen, C. Q., 000. Inversing chlorophyll concentration o sea water using artiicial neural network. Chinese science bulletin, 45(7), pp (in Chinese). REFERENCES: Bricaud, A., and Morel, A., 986. Light attenuation and scattering by phytoplankton cells: a theoretical modeling. Appl. Opt., 5, pp Campbell, J. W., O'Reilly, J. E., 988. Role o satellites in estimating primary productivity on the northwest Atlantic continental shel. Cont. Shel Res., 8, pp Claudia, G., Monica, P., 00. Detecting chlorophyll, Secchi disk depth and surace temperature in a sub-alpine lake using Landsat imagery. The Science o the Total Environment, 68, pp Chen, C. Q., Shi, P., Mao, Q. W., 996. Study on modeling chlorophyll concentration o surace coastal water using TM data. Environmental Science and Technology, (3), pp (in Chinese). Cheng, S. T., Kuang, C., Wang, J. P, 00. Probe into the Water Color Remote Sensing Theoretical Model. Journal o Tsinghua University, 4(8), pp (in Chinese). Cong, S., 998. Neural network theory and application on MATLAB toolbox. The university o science and technology o china press, Anhui,, pp (in Chinese). Ekstrand, S., 99. Landsat TM based quantiication o chlorophyll-a algae bloom in coastal waters. Int. J. Remote Sens., 0, pp Kuang, C., Wang, J. P., Cheng, S. T., etc., 00. The Inversing Study o Water Color Remote Sensing in Sea Area o Macau. In: The ourth symposium on environmental and city development. Beijing, China, pp (in Chinese). Morel, A., Prieur, L., 977. Analysis o variations in ocean color. Limnol. Oceanogr.,, pp Richter, R. A., 990. Fast atmospheric correction algorithm applied to Landsat TM Images. Int. J. Remote Sens.,, pp Richter, R. A., 996. Spatially adaptive ast atmospheric correction algorithm. Int. J. Remote Sens., 7, pp. 0-4.

OCEAN COLOR PRODUCTS RETRIEVAL AND VALIDATION AROUND CHINA COAST WITH MODIS

OCEAN COLOR PRODUCTS RETRIEVAL AND VALIDATION AROUND CHINA COAST WITH MODIS L. Sun a, M. Guo b, X. Wang b a National Satellite Meteorological Center, China Meteorological Administration, No.46 South Avenue