Estimating Rice Yield by HJ-1A Satellite Images

1 1, 2 1 LI Wei-guo , LI Hua , ZHAO Li-hua (1Institute of Economy and Information, Jiangsu Academy of Agricultural Sciences, Nanjing 210014, China; 2Anhui Agricultural University, Hefei 230036, China)

Abstract: Being taken Xuyi County, Jinhu County and Hongze County in Jiangsu Province, China as examples, monitoring and forecasting of rice production were carried out by using HJ-1A satellite remote sensing images. The handhold GPS machines were used to measure the geographical position and some other information of these samples such as areas shapes. The GPS data and the interpretation mark were used to correct HJ-1 image, assist human-computer interactive interpretation, and other operations. The test data had been participated in the whole classification process. The accuracy of interpreted information on rice planting area was more than 90%. By using the leaf area index from the normalized difference vegetation index inversion, the biomass from the ratio vegetation index inversion, and combined with the rice yield estimation model, the rice yield was estimated. Further the thematic map of rice production classification was made based on the rice yield data. According to the comparison results between measured and fitted values of yields and areas of sample sites, the accuracy of the yield estimation was more than 85%. The results suggest that HJ-A/B images could basically meet the demand of rice growth monitoring and yield forecasting, and could be widely applied to rice production monitoring. Key Words: rice; yield; satellite remote sensing images; estimation model

China is one of the biggest rice cultivation estimation. In recent years, some domestic researchers countries. Monitoring and forecasting rice production have made some crop yield estimation research based timely and accurately is very significance for on our satellites data and achieved preferable results adjusting planting structure and making food policy. (Qi et al, 2008; Qin et al, 2009). However, there are In the early 1990s, Chinese Academy of Sciences and few reports about rice yield estimation (Yang et al, other units, organized by the Ministry of Science and 2009). Taking the main rice producing region of Technology in China, used Landsat/TM and Jiangsu Province in China as an experimental area and NOAA/AVHRR images data to estimate planting combining with rice yield estimation model, we used areas and yields of rice, wheat and corn in main HJ-A/B images and conducted related exploratory grain-producing area. The estimation of planting area research on remote sensing yield estimation to exactly and yield of rice was conducted in Hubei and Jiangsu monitor and forecast rice planting areas and yields, Provinces, China and the prediction accuracy was which provided information for rice quantitative above 85%. The main methods and technology system management and macro-control. of crop yield assessment were preliminarily mastered by using of remote sensing data (Xia et al, 1996; Chen MATERIALS AND METHODS et al, 1997; Lobell et al, 2003). As many kinds of satellites such as land resources satellite, environmental Research regional overviews mitigation satellite, meteorological ocean satellite and other series have been successfully launched, China Being taken Xuyi County, Jinhu County and has made great progress in the sensor field of Hongze County in Jiangsu Province, China as study multi-spectral and even hyperspectral. Using our own area (Fig. 1), these regions locate at 118°12'E– satellites data to carry out related rice yield estimation 119°36'E, 32°43'N–36°06'N with temperate zone studies have become an inevitable trend of crop yield monsoon climate and four distinct seasons. It has an annual frost-free period of about 204 d, an average annual temperature of 14°C, an average annual Received: 12 August 2010; Accepted: 30 November 2010 Corresponding author: LI Wei-guo ([email protected]) precipitation of 940 mm and average annual sunshine Rice Science, Vol. 18, No. 2, 2011

was good. Geometric correction was first roughly made using 1:100 000 topographic maps, and then accurately carried out by the GPS route recording and ground truth data to make sure the alignment error within one pixel. According to the experience linear transformation, the atmospheric radiation correction and reflection conversion were obtained by using the actual reflectivity of ground calibration body and the corresponding primitive DN value of the satellite images.

Fig. 1. The distribution of GPS sample sites in study area. Remote sensing yield estimation model of rice

In the sun radiation energy only the visible part ranged from 2130 to 2430 h, which are better climate (350–700nm), accounting for about 47% to 48%, can and soil conditions. In sync with satellite transit time, be used for photosynthesis by plants. The ability of we established 28 sample points for the difference rice daily conversion from visible light into organic GPS fixed-point investigation and sample that included light is called solar contract efficiency (also called leaf area index and biomass. For determine dry matter, daily increase in above-ground biomass weight, sampling plants were dried at 105°C for 20 min, and DABW). Daily increase in above-ground biomass then at 75°C for constant weight. The leaf area index weight calculation referred to Gao et al (1992) was measured by the gravity method (Li and Li, 2010). simulation algorithm, the indication is the type: The grain yield was got by field surveys, which took samples by 50 m×50 m type frame according to the B ⎛ 1+ D ⎞ ΔDABWi = × Ln⎜ ⎟× DL× δ, field block diagonal line 5 points, grains in 1 m2 at K × A ⎝1+ D × Exp()− K × LAIi ⎠ 2 every sampling point with totally 5 m were collected. D = A× 0.47 × (1− α)× Qi / DL , (1) 2 They were naturally air-dried (approximate 13% water Where ΔDABWi [kg/(hm ·d)] is the i-th day of the content) and weighed. The meteorological data were daily increase in above-ground biomass weight, K is provided by the local meteorological department in the group extinction coefficient; LAIi is the i-th day of the three counties. the leaf area index; D is the middle variable; α is the

rice group reflection (%); Qi is the daily global solar Data source radiation (MJ/m2); B and A is the experiment The satellite in this study is China environment coefficients, the value is 22 and 4.5 respectively; δ is disaster reduction satellite, the HJ-1 satellite for short, the conversion coefficient of CH2O and CO2 with 0.68 which includes A star and B star with an orbital value; DL is the day length (h); and LAIi at the altitude of 650 km. CCD (Charge coupled device) heading stage can be computed by the model (Y=2. 1.3655NDVI camera covers the earth every 4 days (HJ-1A and 8339×e (Li et al, 2008)). NDVI is the HJ-1B satellite network revisit cycle is 2 days over the normalized difference vegetation index, NDVI= (RNIR– global), spectrum scope covers blue light (0.43–0.52 RRED)/(RNIR+RRED), where RNIR is the near infrared μm), green light (0.52–0.60 μm), red light (0.63–6.90 reflectance (the fourth band of HJ-1A satellite), and μm) and near-infrared light (0.76–0.9 μm), the RRED is the red reflectance (the third band of HJ-1A subsatellite resolution is 30 m, the breadth of single satellite ). CCD camera is 360 km (the breadth of two cameras is Daily increase in above-ground biomass weight 710 km). removing the plant growth respiration and its The HJ-1A satellite transit time was on 26 maintenance consumption is the plant daily net August, 2009, when rice was at the heading stage. It assimilation (also called the above-ground biomass was sunny and cloudless, so the satellite image quality weight, ABW). Its algorithm is as follows: LI Wei-guo, et al. Estimating Rice Yield by HJ-1A Satellite Images

Table 1. Parameters of rice materials. Where i is the days from sowing to maturity (d), Seeding rate Growth Harvest which equals to growth duration, and β is the ratio of Rice material 2 K (kg/hm ) duration (d) index root to shoot the maturity stage, generally between Liangyou 6 150 130 0.41 0.63 Liangyoupeijiu 150 140 0.43 0.67 0.05 and 0.08, the value is 0.06 in the model. The Xudao 3 150 152 0.45 0.69 parameter information of rice yield estimation model is listed in Table 1. ΔABWi = (ΔDABW i-RGi-RMi) × min(TFi, NF ) (2) Where RG and RM are the photosynthesis i i RESULTS assimilation amount [kg/(hm2·d)]. They are growth breath consumption [kg/(hm2·d)] and maintenance Forecast and confirmation of leaf area index and breath consumption [kg/(hm2·d)] of the i-th day biomass respectively. TFi and NF are influence factor of above- ground biomass accumulation from air temperature The values of RVI and NDVI and the information and soil nitrogen respectively. Its algorithm was of the 28 sample points were extracted through the described by Li et al (2008). ERDAS and ENVI softwares, and the LAI and During the rice growing period, the above- biomass in each sample point were calculated by an ground biomass weight is the sum of dry matter inversion model. The rice LAI from inversion was low, accumulation. ABWi is the total above-ground dry with the root mean square error (RMSE) of 0.38 (Fig. biomass accumulation (kg/hm2) of the i-th day from 2-A), which indicated that the data was comparatively the first emergence day to harvest. ΔABWi is the consistent with relative error ranged from 1.0% to increment on the i-th day of the above-ground dry 10.3%, and the average error of 4.72%. It is showed biomass weight [kg/(hm2·d)] and described as follows: that the LAI inversion model could be used to

n ABW i = ∑ ΔABW i (3) estimate the current rice LAI because of reliable i=1 results. In Fig. 2-B, the actual biomass values were a Where n is crop growth duration (d); ABW1 (the bit higher than those from the model inversion with first emergence day of above-ground biomass) is half the RMSE of 464.14 kg/hm2, the relative error ranged of the weight of sowing seeds; and ABWi at the rice from 1.57% to 7.57%, and the average error of 4.96%, heading stage can be got through the inversion model which stated the biomass inversion model could be (y=741.76×RVI+4253.2). RVI is the ratio vegetation used to estimate current rice biomass. index, RVI=RNIR/RRED. Rice yield refers to the product of rice grain or Rice sampling point data and confirmation of brown rice, which is the product of above-ground estimation model precision biomass at the maturity stage and harvest index (HI). According to the design features of rice yield The algorithm is as follows: estimation model, rice yield in the area under study i (4) Y = [ ABW /(1 + β)] × HI was predicted based on rice variety parameters,

Fig. 2. Comparison of fitted rice leaf area index, biomass and yield with measured values. Rice Science, Vol. 18, No. 2, 2011

Table 2. Data information of sampling sites at the rice heading stage. Sampling Leaf area index Biomass (kg/hm2) Measured yield Fitted yield NDVI RVI site Measured Fitted Measured Fitted (kg/hm2) (kg/hm2) XY01 0.51 3.14 6.13 5.69 7 121 6582 6915 6 450 XY02 0.78 8.22 8.02 8.26 9 987 10 350 9 886 10 296 XY03 0.57 3.69 6.89 6.18 7 531 6 990 7 560 6 951 XY04 0.77 7.8 8.46 8.14 10 503 10 039 10 868 10 016 XY05 0.64 4.63 7.23 6.81 8 269 7 688 8 300 7 724 XY06 0.54 3.34 6.47 5.93 8 355 7 858 8 235 7 502 XY07 0.76 7.17 8.47 7.95 10 335 9 720 10 350 9 701 XY08 0.75 6.92 7.59 7.87 9 065 9 386 9 160 9 419 XY09 0.74 6.73 7.52 7.8 8 703 9 245 8 650 9 285 XY10 0.74 6.67 8.13 7.78 9 652 9 200 9 730 9 243 XY11 0.69 5.46 6.78 7.28 7 965 8 303 7 890 8 372 XY12 0.72 6.08 7.35 7.56 8 543 8 763 8 550 8 828 JH13 0.76 7.5 7.98 8.06 9 664 9 816 9 700 9 817 JH14 0.7 5.75 7.68 7.41 8 932 8 518 8 820 8 585 JH15 0.75 7.09 7.53 7.93 9 012 9 512 9 280 9 537 JH16 0.66 4.88 7.42 6.98 8 446 7 873 8 540 7 931 JH17 0.75 6.9 7.36 7.86 8 897 9 371 8 900 9 403 JH18 0.67 5 7.49 7.05 8 556 7 962 8 440 8 025 JH19 0.73 6.28 8.22 7.87 9 433 8 911 9 500 9 062 HZ20 0.75 6.83 8.36 8.08 9 859 9 319 9 940 9 452 HZ21 0.71 6.67 7.15 7.48 8 724 9 201 8 800 9 125 HZ22 0.79 8.44 8.06 8.32 10 068 10 514 9 960 10 443 HZ23 0.66 4.99 7.26 6.98 8 356 7 955 8 450 7 992 HZ24 0.71 6.58 7.64 7.48 9 553 9 134 9 650 9 075 HZ25 0.7 5.83 6.98 7.36 8 123 8 578 8 250 8 610 HZ26 0.78 8.15 8.45 8.23 9 899 10 299 10 565 10 246 HZ27 0.75 6.95 8.19 7.89 10 423 10 150 9 900 10 000 HZ28 0.67 4.69 6.79 7.07 7 530 7 732 7 986 7 861 For the codes of sample sites, XY, JH and HZ represent Xuyi County, Jinhu County, and Hongze County, Jiangsu Province, respectively. The rice cultivars planted in the three counties were Liangyou 6, Liangyoupeijiu and Xudao 3, respectively. NDVI, Normalized difference vegetation index; RVI, Ratio vegetation index. meteorological data and remote sensing information. documents by using the existing administrative The rice actual yield was between 6 915–10 868 boundary vector, cut out the experimental region from kg/hm2, with the average of 8 929.96 kg/hm2; and the HJ-A images, and generated the interpretation base estimated yield was between 6 450–10 443 kg/hm2, image choosing 4, 2 and 1 bands combination. After with the average of 8 859.96 kg/hm2 (Table 2). the classification and interpretation process, rice Fig. 2-C demonstrated the 1:1 relations of planting area of the sample cities and counties in 2009 different sampling points between rice actual yields was 131 022.27 hm2, while the actual area was 126 and the estimation ones. The standard deviation 200 hm2, so the interpretation accuracy was over 90%, 2 RMSE of yield was 482.5 kg/hm , and the relative which proved the result was much more reliable. error was in a range of 1.01%–8.90%, with an average GIS software was used to get the regional of 4.90%. The model accuracy test, between the production information map based on linear estimation values and the actual values, showed that transformation between sample point yield data from the precision was over 85%, which indicated that it rice yield estimation model and NDVI image. was possible to estimate the regional rice yield According to the rice single-yield, the rice fields were combining biomass and leaf area index at the heading divided into four types: higher-yielding fields (Rice-I, stage through rice yield estimation model. ≥ 9 750 kg/hm2), high-yielding fields (Rice-II, 8 250–9 750 kg/hm2), medium-yielding fields (Rice-III, Extraction of rice planting area and construction 6 750–8 250 kg/hm2) and low-yielding fields of levels of production forecast figure (Rice-IV<6 750 kg/hm2). Monitoring and stage Accurate extraction of rice planting area by prediction thematic map of rice yield in the region was remote sensing is very important preparatory work for gained by overlay the actual sample point rice yield rice production estimation. We made the county AOI data (Fig. 3). LI Wei-guo, et al. Estimating Rice Yield by HJ-1A Satellite Images

16.57%, and 14.30% of the total rice area, respectively. The rice was at the heading stage when sampled, which was a key stage for rice production. If some managements can be applied in the low-yielding fields, such as fertilization and reasonable irrigation, it would increase rice yield and make the low-yielding fields develop to medium-yielding fields. Similarly, some scientific managements and reasonable regulation based on soil properties could be conducted in the medium-yielding fields to reach high-yield.

DISCUSSION Fig. 3. Gradation monitoring and forecasting chart of rice yield. Actual rice planting area data was provided by the local agricultural management department. Rice-I, II, III, and IV represent Xuyi County, Jinhu County and Hongze County, the area of field, in which the yield is higher than 9 750 kg/hm2, 8 in the middle of Jiangsu Province, China, have the 250–9 750 kg/hm2, 6 750–8 250 kg/hm2, and lower than 6 750 kg/hm2, respectively. undulating topography and complex planting structure. The lower spatial resolution satellites (for example, Xuyi County has many hills and mountains, so NOAA, MODIS and so on) cannot satisfy the the most high-yielding fields and high-yielding fields accuracy requirement of crop growing and yield only accounts for 65.9%. However, the natural estimation, particularly for precise monitoring in south environment, climate, moisture content, topography plant area, the regions with complex terrain and and so on in Hongze County, is suitable for rice diverse farming system. This research selected HJ-1A planting, and the most high-yielding fields and satellite images with 30 m spatial resolution, which high-yielding fields were quite more, accounting for can satisfy the accuracy requirement for monitoring 78.3%. crop growing in the study area. Moreover, HJ-1 Table 3 is the area distribution of various rice satellite, after combining A star with B star, can revisit yield levels, from which we can conclude that the rice the same place in 2 days and much better satisfied the extraction area precision in each county was above time precision for rice long-term dynamic monitoring. 85% and the area extraction result was comparatively On the rice yield remote sensing monitoring and reliable. In different counties under study, the rice forecasting, most previous research always focused on yield grade area was different, Xuyi County possessed analyzing the relationship between spectral more low-yielding fields, and Hongze County has a information of images with rice LAI and biomass or high proportion of the most high-yielding fields and yield by establishing a regression model (Liu et al, high-yielding fields. In the three counties, the areas of 2004; Tang et al, 2004; Cheng et al, 2006). Although the higher-yielding fields, the high-yielding fields, the the yield estimation model is simple, it is strongly medium-yielding, and the low-yielding fields were empirical and less universal. Remote sensing images 16993.20 hm2, 73587.98 hm2, 21708.20 hm2 and can obtain rice instantaneous growing information at a 18732.90 hm2, which occupied 12.97%, 56.16%, certain growth stage, but it always has a large

Table 3. The distribution of rice area at different yield levels after classification. Planting area Interpretation area Interpretation Rice-I Rice-II Rice-III Rice-IV County (hm2) (hm2) accuracy (%) (hm2) (hm2) (hm2) (hm2) Xuyi 62 700 60 055.05 95.78 6 665.76 32 953.98 9 625.40 10 809.91 Jinhu 36 000 39 895.86 90.23 4 764.40 21 852.90 8 690.02 4 588.54 Hongze 27 500 31 071.36 88.51 5 563.04 18 781.10 3 392.78 3 334.45 Total 126 200 131 022.27 91.51 16 993.20 73 587.98 21 708.20 18 732.90 Actual rice planting area data was provided by the local agricultural management department. Rice-I, II, III, and IV represent the area of field, in which the yield is higher than 9 750 kg/hm2, 8 250–9 750 kg/hm2, 6 750–8 250 kg/hm2, and lower than 6 750 kg/hm2, respectively. Rice Science, Vol. 18, No. 2, 2011 deviation to predict yield at the maturity stage. 12(4): 363–369. (in Chinese with English abstract) Because the climate conditions (temperature, light, Chen J Y, PAN Delu, Mao Z H. 2006. Optimum segmentation of water status etc.) were constantly changeable at the simple objects in high-resolution remote sensing imagery in coastal areas. Sci China Ser D: Earth Sci, 49(11): 1195–1203. prediction period, which had a great influence on rice Cheng Q. 2006. Models for rice yield estimation using remote production. Combining the continuous and dynamic sensing data of MOD13. Trans Chin Soc Agric Engineering, feature in the crop growth model with instantaneous 22(3): 79–83. (in Chinese with English abstract) and wide-area advantage in the remote sensing, it can Gao L Z, Jin Z Q, Huang Y. 1992. Rice Cultivation Computer play a ‘point’ and ‘surface’ complementary effect. Simulation Optimization Decision System. Beijing: China This research used rice yield estimation model Agricultural Science and Technology Publishing House: 29–33. (in Chinese) which coupled with remote sensing data and the Li W G. 2007. A classification of rice yield based on TM image growth model, integrated remote sensing inversion data and estimating yield model. Jiangsu Agric Sci, (4): 12–13. information and made yield estimation through (in Chinese) assimilation. The estimated yield and actual values Li W G, Li Z J. 2010. Study on estimating winter wheat yield by had a high consistency with more than 85% yield CBERS satellite images. J Trit Crop, 31(2): 172–175. (in prediction. The results indicated that the estimation Chinese with English abstract) model can be effectively used to estimate the rice Li W G, Wang J H, Zhao C J, Liu L Y, Tong Q X. 2008. Estimating rice yield based on quantitative remote sensing yield, which had a good theoretical value and practical inversion and growth model coupling. Trans Chin Soc Agric prospect because of improved yield estimation Engineering, 24: 128–131. (in Chinese with English abstract) accuracy and strengthened remote sensing yield Liu L Y, Wang J H, Huang W J, Zhao C J, Zhang B, Tong Q X. estimation mechanism. In addition, the rice yield 2004. Improving winter wheat yield prediction by novel remote sensing classification forecasting map in this spectral index. Transa Chin Soc Agric Engineering, 20(1): study had a better practicability for primary 172–175. (in Chinese with English abstract) Lobell D B, Asner G P, Ortiz-Monasterio J I, Benning T L. 2003. agricultural technicians to guide the field production Remote sensing of regional crop production in the Yaqui management and planting division. Valley, Mexico: Estimates and uncertainties. Agric, Ecosys It still exists mixed pixel problem in HJ-A/B Environ, 94: 205–220. satellite images. Hereafter, if we can unify the high Qi L, Zhao C J, Li C J, Liu L Y, Tan C W, Huang W J. 2008. resolution images (for example, SPOT, QUICKBIRD Accuracy of winter wheat identification based on and so on) and use image fusion technology to display multi-temporal CBERS-02 images. Chin J Appl Ecol, 19(10): the advantage of multiple source and multi-temporal 2201–2208. Qin Y W, Zhao G X, Jiang S Q, Cheng J N, Meng Y, Li B H, Xu G remote sensing images (Chen et al, 2006; Zhang et al, C, Han J G. 2009. Winter wheat yie1d estimation based on 2008), the classification and interpretation precision high and moderate resolution remote sensing data at county can be greatly improved, and the rice growing trend 1evel.Transa Chin Soc Agric Engineering, 25(7): l18–l23. (in graduation monitoring accuracy can be enhanced. Chinese with Eng1ish abstract) Moreover, we analyzed the new rice yield remote Tang Y L, Wang J H, Huang J F, Wang R C. 2004. Yield sensing estimation model with some innovation. estimation by hyperserctral data of rice canopies in mature stages. Acta Agron Sin, 30(8): 780–785. (in Chinese with However, some individual sample points had larger English abstract) deviation, which might be affected by model Xia D S, Li H. 1996. The status quo of remote sensing application parameters or environmental factors. In the future, we for natural disaster in some countries. Remote Sensing Land & need to validate it in a wider area to further clarify the Resour, 29(3): 1–8. (in Chinese with English abstract) boundary conditions of this yield estimation model. Yang W D, Song Y T, Song X Y, Ding G W. 2009. Winter wheat yield estimating based on 3S integration and field REFERENCES measurement. Trans Chin Soc Agric Engineering, 25(2): 131–135. (in Chinese with English abstract) Zhang H, Yao X G, Zhang X B, Zhu L L, Ye S T, Zheng K F, Hu Chen S B, Sun J L. 1997. Key technical links and solution ways of W Q. 2008. Measurement of rice leaf chlorophyll and seed setting up a working system for yield estimations of the main nitrogen contents by using multi-spectral imagine. Chin J Rice crops of China by satellite remote sensing. J Nat Resour, Sci, 22(5): 555–558. (in Chinese with English abstract)