| Rice is the most important food crop in China. The change of cultivated area, pros and cons of growth and the richness and deficiency of yield have been always highly emphasized by all levels of government and been the important basis in national or regional grain policy decision and economic development plan. Therefore, understanding the factors that influence the rice yield, monitoring the planting area and estimating the rice yield are of great importance to the food securityRemote sensing data, due to its timely update, full coverage and objective accuracy, has been widely applied in real-time monitoring of rice growth and rapid yield estimation. In recent years,3S technology typically remote sensing technology, has made some achievements on the monitoring of rice area and yield estimation. But in complex terrain regions affected by scattered paddy field, climate change and different cropping systems, large scale yield estimation based on remote sensing data still has some limitations and the accuracy is not high.Chongqing, located in the southwest of China, is characterized by hilly and mountains, with complex and rugged terrain and significant regional elevation difference. The distribution of paddy field is dispersed and map spots of paddy are broken. Besides, climate is humid with larger cloudiness covered and heavy fog in the region. Remote-sensing image is easily disturbed by cloud and mist, further increasing the difficulty of yield estimation from remote sensing data in the study area. At present, there are few researches on yield estimation from remote sensing data in this area. As the latest direct-controlled municipality of the central government, urbanization level in Chongqing is not high. Agriculture is the basis of economic development and rice production typically plays the most important role in local agriculture. Therefore, understanding the characteristics of local rice distribution, seeking main factors that influence rice yield and estimating yield is practically meaningful for the development of agricultural and economic in Chongqing. It will also provide the references for the improvement of rice yield estimation precision and industrialization of yield estimation in other hilly mountains.In this study, Chongqing city was chosen as the study area and Savizky-Golay filtering method was utilized to deal with the MODIS-NDVI (Moderate Resolution Imaging Spectroradiometer-Normalized Difference Vegetation Index) data from 2003 to 2012 to wipe off the noise and establish remote sensing estimation model at different regional scale. Based on the analysis about ecological environment data and the rice yield data of rice-growing area in Chongqing, this research determines the main factors that influence the estimation of rice yield, establishes remote sensing estimation model of rice yield at different regional scales and adds the main factors, improving the model accuracy. The main methods and results are as follows:(1) Characteristics of topography distribution in Chongqing rice-growing areaFour ecological zones including the west of Chongqing (hereafter YX), the centre of Chongqing (YZ), the northeast of Chongqing (YDB) and the southeast of Chongqing (YDN) were delineated by the terrain, climate, soil, and rice cropping system as well as the integrity of administrative district into consideration. In order to investigate the terrain distribution characteristics, four terrain attributes including elevation, slope, aspect and slope positions were taken into account in rice planting areas. The results indicate that different ecological regions present different terrain distribution characteristics. The paddy fields in YX account for 48.7% of the study area The distribution of these paddy fields concentrated in aspects of east and southeast and the flat slope position under 6 °, with an altitude from 200 to 400 m. Paddy area in YZ accounts for 35.1% of the study area. The paddy is mainly distributed at an altitude of 300-500m, the aspect of northwest and middle slope position, with a slope under 15°. Paddy area in YDN accounts for 7.4% of the whole city. The paddy is mainly distributed at an altitude of 300- 900 m and the slope of 6-25°. Specially, paddy with an elevation of 300-400 m and the slope of 6-15° is the most widely distributed. Most of them mainly focus in aspects of southeast and west and middle slope position. Paddy area in YDB accounts for 8.8% of the whole study area. The paddy is mainly distributed at the elevation under 1000m and the slope of 6-25°. Specially, paddies with an elevation of 600-800 m and the slope of 6-15° are the most widely distributed. Most of them mainly focus in aspects of south and middle slope position.(2) Characteristics of climate in Chongqing rice-growing areaThe conventional statistical method is applied to analyze the mean temperature, sunshine, precipitation and their annual changes and the period of rice growth (April-September) from 1985 to 2012. In recent 30 years, the annual average temperature was 17.52 ℃, sunshine hours were 1163.3 h, and amount of precipitation was about 1105 mm. During the rice growth period, the annual average temperature was 23.75 ℃, sunshine hours were 838.46 h, and precipitation was about 850 mm. Both annual precipitation means and rice growth period precipitation, the variations exhibit the largest with 11.41% and 15.78% respectively.Those calculated results acquired from the method of Mann-Kendall test demonstrated that in recent 30 years, the annual mean temperature of Chongqing has significantly rising trend at 0.05 levels. Although the rainfall has declined in a way the tendency is not obvious. Also the rising trend of sunshine duration is not apparent. On the other hand, the annual mean temperature for rice growth period has significantly rising trend at 0.01 levels. In spite of the trend of precipitation and sunshine duration was down, but the decline is not distinct.In this study, the method of thin plate spline interpolation was applied to simulate the spatial structure of each meteorological factor during the period of rice growth. The results show that the annual average temperature of rice growing period in the low-altitude region, is higher. With the increase of altitude, the average temperature is decreasing. Sunlight in the rice key growing period is the strongest in YDB. Sunlight in YZ takes the second place, followed by YX. Sunlight in YDN is the least. Precipitation of the rice key growing period increases gradually over time, reaches the maximum in June and July, and then shows a downward trend. Precipitation mainly concentrates in YDN and YZ, followed by YDB, and it is the least time in YX in April and May. Precipitation in YDN in June is significantly higher than other areas. Precipitation increases in YDB from July to September, the trend of which in YDB is higher than that in YDN.(3) Spatio-temporal variation and affecting factors of rice yield.A linear model was developed to analyze the temporal variability of rice yield and production. The results show that rice yield from 1985 to 2012 has a significantly upward trend, but the trend for rice planting area is significantly down over time. Rice production exhibits slight decline trend, however, it is not significant. In the long time sequence, the results of correlation analysis indicate that the precipitation was positively correlated with rice yield (P<0.1). Results of path analysis showed that temperature and sunshine have a crucially direct effect on rice yield, but the influences on rice yield have been weakened by other meteorological factors. In conclusion, temperature and sunshine are the most critical factors for rice yield when rainfall is adequate (irrigation is guaranteed)The statistical results on rice yields from different ecological countries of study area from 2003 to 2012 show that the diminishing order of rice yield in the four ecological zones is YX, YZ, YDN and YDB. The diminishing order of rice planting area is YX, YZ, YDB and YDN. And the diminishing order of rice production is YX, YZ, YDB and YDN.In order to understand the factors that affect the rice yield estimation, in this study, yield, terrain and climate data derived from various townships in 2008 were used as data source and the model of classification and regression tree (CART) was utilized to calculate the relative importance of various ecological factors on rice yield. The results indicate that in the areas where the altitude is relatively lower with small elevation difference like YX, terrain is the main factor to affect yield. In the areas where altitude is relatively higher with strong elevation difference like YDN, YZ and YDB, climate is the decisive factor. In YDN, sunlight is the major climate factor whereas in YX, YZ and YDB mean temperature is the key climate factor. (4) Spatio-temporal distribution of NDVISpatial and temporal variability of NDVI was explored during the period of 2003-2012. The annual NDVI value fluctuates from 0.54 to 0.59. The NDVI values of ten years show an upward trend of vegetation coverage, which shows that overall situation of vegetation coverage in Chongqing is good. The characteristics of NDVI space distribution are characterized by:YDB>YDN>YZ>YX.It also studies the differences of rice growth cycle in different ecological zones and different elevation. Studies have shown that before the NDVI reaching the maximum, in the same time phase, the NDVI in low altitude area is generally higher than in high altitude area in regions such as YX, YZ and YDB, which confirms that growth time of rice in low altitude area is earlier than that in high altitude area. NDVI curves of rice growth in the YDN and three other ecological zones have obvious difference. The time of NDVI at different altitudes in YDN reaching the maximum is a one-phase time later than other ecological zones. This result proved that compared with the other ecological zones under the same elevation, the rice field growth period in YDN is delayed.The relationship between NDVI of Rice and climate in the study area is shown as: NDVI from May and June was significantly correlated with sunshine hours and the average temperature, however, it was negatively correlated with precipitation; NDVI from July and August was negatively correlated with sunshine hours and the average temperature but significantly correlated with precipitation; In April and September, only sunshine had the significantly positive correlation with rice NDVI.In order to further study understand the relationship between NDVI and climate, this research utilized drought index model to simulate space distribution of drought in 2006, and used the drought classification map to extract pure pixel NDVI values of paddy field. Paddy pure pixel NDVI curves indicate that if drought in one area is more serious, NDVI value of paddy field falls faster. NDVI declining trend of paddy field can explain the response of NDVI to drought.(5) Estimation of rice yield at county scaleThe topography in Chongqing is mainly hilly and mountainous. The distribution of paddy field is dispersed and map spots of paddy are broken. There are a lot of mixed pixels. For the similarities of climate, soil, and management level in the same county, NDVI changes of rice in the area also have similarity. This research uses average NDVI of 277 priority paddy field pure pixels, which derived from 2003 to 2011 to replace paddy field NDVI from the counties where these pure pixels are located.In order to build rice production forecasting model, the average NDVI is utilized to multiply the results of rice planting area (ANDVI). And the specific procedure is depicted as follows:1) Establish multi-perspective rice yield estimation model according to different areas and ecological zoning regions 2) Establish the rice yield estimation model with different algorithms including the stepwise regression (SRM), BP neural network (BPNN), support vector machine (SVM) and random forests (RF).3): Establish multi-factors rice yield estimation model The model is established with ANDVI and rice yield, and the compound model is setup in line with ANDVI, meteorological data and rice yield. On this basis, the optimal estimation model is adopted to forecast the rice yield in 2012. The results show that rice yield estimation model of partition is superior to regional model inordinately. Although both the neural network model and support vector machine model have shown a high fitting precision, the model of support vector machine has better result. And that the accuracy of composite estimation model which ANDVI together with climate is higher than simple ANDVI.In order to obtain the best time phase for rice yield estimation, various ANDVIs at different time phases during the rice growth was putted as the input parameter and the optimal rice yield estimation model of regional and partition was applied to point out some phenomenon through the comparison of fitting precision. The results show that the optimal estimation phase for rice output is 185 at county scale. Additionally, for different ecological regionalization the best estimation phase was different, for YX the best phase is 185, for YZ and YDB is 201, and for YDN is 217. Because climate and topography are different in different ecological zones and the growth period of rice has a certain difference, the best phase of rice estimating is different. But they are all in the vegetative and reproductive growth stage of rice.(6) Estimation of rice yield at township scaleIn order to explore the estimation ability of NDVI data with 250m resolution at township-level, the yield data at township-level of Chongqing in 2008 was utilized, region and ecological zoning are also applied to the remote sensing estimation of rice production in township scale in 2008. Remote sensing models including SRM, BPNN, SVM and RF are established with ANDVI and rice production at township scale. Remote sensing and ecological factor complex model including SRM, BPNN, SVM and RF are also built with ANDVI, ecological factors and rice production at township scale. Moreover, the optimal fitting model of rice production is used to predict the output in 2009. Results indicate that remote sensing model based on ecological zoning is more adaptive than others, and the model including terrain and meteorological factors is the optimal. This also indicates temperature, precipitation, sunshine climatic conditions and terrain conditions have important effect on rice yield. Rice production estimation method of BPNN shows the optimal fitting performance.various ANDVIs at different time phases during the rice growth in 2008, was taken as the inputs, as well as the optimal rice yield estimation model of regional and partition was applied to determine diversities of the best estimate phase in the different regions. It turned out that in 2008, the optimal estimation phase for rice output was 169 in township scale. Additionally, for different ecological regionalization the best estimation phase was different, for west the best phase is 169, for central and northwest is 185, and for southeast is 201. Compared with the results fitted with the county data, it is a phase ahead of time. However, it reflects the same regularity, namely; similar to YDB the optimal predicted time phase in YZ have a time-phase delay compared to YX. And the optimal predicted time phase in YDN has a two-time-phase later than YX. On the other side, the best prediction period for all ecological zones is still at the second stage of growth that the vegetative and reproductive.(7) Estimation of rice yield at village scaleTo study the applicability of NDVI with 250m resolution on yield estimation, and investigate the influence of soil factors on the yield estimation model, village-level data of 2007 with yield, terrain and soil was collected from Hechuan, Chongqing. Since climate change is not obvious in small regions, the impacts of climate change on rice yield have not been taken into consideration. The results show that soil factors have stronger effect on rice yield than that from terrain factors. The relative important of alkali-hydrolysable nitrogen, soil potassium and PH has reached up to 50% whereas only slope in terrain factors is more than 50%.Bilinear interpolation was applied to resample NDVI data into image maps with 30m. And the land use status data with 1:1 million of Hechuan was used to extract the paddy fields distribution. Four methods including stepwise regression, neural network, support vector machine regression and random forest were utilized to establish the rice yield estimation model and the composite model that ANDVI with ecological factors at the level of village. The optimal model in 2007 is used to forecast rice output in 2008. The results show that composite model with ANDVI and ecological factors which based on neural network model has the optimal fitting performance, with the determination coefficient is 0.732and the relative mean square error is 25.8%. The best time phase for Hechuan from the optimal estimation model for rice yield estimation is 185. |
PDF Full Text Request