EVALUATION METHOD OF WATER QUALITY FOR RIVER BASED ON MULTI-SPECTRAL REMOTE SENSING DATA

With the rapid development of the regional economy, water pollution has gradually become an environmental problem that cannot be ignored. As an important water source in central China, the Han River should strengthen water quality monitoring and management in order to ensure the sustainable development of watershed and related areas. Taking typical sections of middle and lower reaches of the Han River as the study area, this paper focuses on rapid river water quality assessment using multispectral remote sensing images. Based on measured water quality data and synchronous spatial high and medium-resolution remote sensing data (multi-spectral data of ZY3 and HJ1A) in 2013, neural network algorithm is used to establish water quality index retrieval model for the study area, and then water quality status is assessed accordingly. The results show that BP neural network retrieval model of water quality index that is established based on multispectral data of ZY3 satellite has higher accuracy and that its assessment results are of high credibility and strong applicability, which can really reflect changes in water quality and better achieve water quality assessment for the study area. In addition, water quality assessment results show that major excessive factors in the study area are total nitrogen and total phosphorus; the polluting type is organic pollution; water quality varies greatly with seasons.


INTRODUCTION
Water quality evaluation is a fundamental link in water environment management and monitoring. Only through water quality monitoring can water quality be reasonably evaluated and targeted water environment management planning and scheme be developed. In terms of water quality evaluation, traditional methods like water sample collection, indicator analysis and grade evaluation can only provide water quality status at the sampling point instead of large area of waters, while large-scale field sampling will consume a large amount of manpower, materials and financial resources. In recent years, with the rapid development of remote sensing technique, more and more researchers carried out fast, continuous and dynamic monitoring on waters by means of remote sensing technique.
Further, this technique has been adopted by lots of domestic and foreign scholars on water quality evaluation (Wu, 2012, Gu, Although remote sensing technique exhibits many advantages in terms of water quality evaluation, current studies mostly adopt medium-resolution remote sensing images. Due to their time advantage (for instance, the HJ-1A/1B developed by China independently has a revisiting cycle of 4 days), water quality status can be monitored in a real-time manner and water quality evaluation can be updated rapidly. Yet, as these images have relatively low resolution, their application to lakes with smaller inland area, narrow rivers or reservoirs are largely limited. On January 9, 2012, a civilian high-resolution stereo mapping satellite "ZY-3", the first one of its kind in China, was successfully launched of which the multispectral data's resolution is 5m and the revisiting cycle is 5 days. Through the satellite, nationwide multispectral images can be obtained in a continuous, stable and rapid manner over a long period of time.
With typical section of middle and lower reaches of Han River as the study area, based on BP neural network algorithm, this paper makes use of the measured water quality monitoring data acquired in summer and autumn in 2013 as well as the multispectral data of the satellites ZY-3 and HJ-1A to establish a water quality parameter retrieval model of the study area, conduct water quality evaluation of the Han River and draw a water quality map of the study area.

Study Area
The middle and lower reaches of Han River is not only an important water source for cities along the River but also serves as a water body with important water environment functions.

MODELS
BP neural network model (back-propagation) is the most common one among neural network classifiers. One of its most important applications is function approximation. It can create any non-linear non-significant function mapping relationship from input to output for the training set and is suitable for the quantitative remote sensing retrieval study of water quality parameters. As a result, BP neural network is used in this paper to establish a retrieval model for concentrations of TN and TP on the basis of multispectral data from HJ-1A and ZY-3. in consideration of fitting speed and accuracy, S-type function is adopted for the neurons in the hidden layer, while a linear function is adopted for the output layer; the number of the neurons in the hidden layer is 6 and the network structure is 4-6-1(for TP, the structure is 2-6-1).

Comparisons Results
Based on the multispectral data of ZY-3 and HJ-1A, the concentration of water quality parameter in the study area is retrieved, and then the water quality evaluation results are obtained from the retrieval results. RE and RMSE are adopted to compare the water quality evaluation accuracies of these two kinds of multispectral data. See details in Figure2.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-7/W3, 2015 36th International Symposium on Remote Sensing of Environment, 11-15 May 2015, Berlin, Germany

Water Quality Evaluation
By means of BP neural network (resilient BP algorithm), based on the multispectral data of ZY-3 and HJ-1A, the spatial distribution map of single factor water quality identification index for TN and TP in the study area is made in the ENVI and MATLAB2013a environment, as shown in Figure 3.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-7/W3, 2015 36th International Symposium on Remote Sensing of Environment, 11-15 May 2015, Berlin, Germany It can be seen from this map that, compared with SFWQII based on measured data (the figures marked on the map reflect the SFWQII obtained from measured data), both these two kinds of multispectral data can get the evaluation results similar to measured data. However, the comparisons indicate that, the spatial distribution map for SFWQII from ZY-3's multispectral data has a higher accuracy, which not only accurately reflect the level of distribution of water quality indicators in the study area (the arrow points out the high value area), but also precisely identify non-water parts (as shown by the circles in Figure 3. (a) and Figure 3. (b)); Yet, as for the spatial distribution map for SFWQII from HJ-1A's multispectral data, although it is able to reflect the overall water quality status of the study area, the levels of distribution of water quality indicators can only be roughly expressed; when we need to know the upstream and downstream statuses of a certain section with high indicator value, the specific evaluation results acquired are less reliable than those results from ZY-3's multispectral data, which is mainly attributed to the low spatial resolution (30m) of these data.

CONCLUSIONS
The water quality evaluation conducted based on remote sensing data supplements the traditional evaluation work. In this paper, the multispectral data from two domestic satellites (ZY-3 and HJ-1A) are used to establish a remote-sensing retrieval model of concentrations of TN and TP through BP neural network (resilient BP algorithm) and conduct water quality evaluations on typical section of middle and lower reaches of the Han River. Comparisons reveal that, ZY-3 s' multispectral data can get more reliable water quality evaluation results and higher-resolution spatial distribution map of these results. Besides, the evaluation results based on ZY-3 s' multispectral images not only reflect the overall water quality status of the study area, but also reveal the upstream and downstream water quality statuses of a certain section with high indicator value and precisely identify non-water parts.