Rice farming is one of the most critical activities in the global agriculture sector, producing a staple food for the majority of the world’s increasing population, which is expected to exceed 9 billion by 2050 and will require approximately 60% more food. Accurate and up-to-date assessments of the spatial distribution of rice cultivation areas are a vital requirement for all stakeholders including policy makers, rice farmers and consumers. Timely assessment with high precision is also crucial for water resource management, market price control and handling of humanitarian food crises [1
To maintain stable rice production, Thailand has recently introduced community-based policies to create regional agricultural plantations. The concept of these policies is to encourage aggregation of small rice fields (farmer-based) into one large (community-based) field. There are three major requirements of this policy: expert field managers, an action plan and site-specific technology [4
In Thailand, where rice farming is commonly practiced in a water-shortage environment, development of management strategies is essential for improving crop production. Under limited water resource conditions, crop models are useful tools for supporting decision-making regarding effective field management. However, such models are complicated and require a large number of parameters. Models also require advanced skills from end-users for model calibration and operation [6
]. For example, the DASST [8
] and APSIM models require many parameters [6
] and the CropSyst, WOFOST and SPAC models are relatively complicated for end-users. These difficulties prevent the development and extensive application of these models [6
Traditionally, crop monitoring and yield estimation are based on ground surveys and reports but this approach is expensive and time-consuming [11
]. As technologies have developed, crop modeling has become increasingly important for providing the necessary information to farmers. Forecasting crop yield is essential for decision-making regarding soil and plant management, crop selection, marketing, storage and transport [14
]. Specifically, forecasting the crop yield is vital for assisting in agricultural planning and decision-making policies. For example, small-scale farmers require information on crop yield to allow them to estimate costs and potential profits before investing their time and money in crop cultivation [15
Farm-scale yield data have become a primary factor in agriculture policy and crop insurance ratings. Nevertheless, researchers often have trouble analyzing these programs because existing farm-scale yield data are sparse, not necessarily broadly representative and might suffer from selection bias [16
]. Crop growth models are useful for estimating growth, development and yield but they need high-quality spatially-distributed input data to derive reliable estimates [17
]. Comparison of the simulated rice yield of small-holder farmers and their different agricultural behaviors would provide necessary information for implementing Thailand’s community-based policy. Agricultural monitoring and management have dramatically benefited from the increased availability of a wide variety of remotely-sensed satellite imagery, ground-sensed data (e.g., weather station networks) and crop models, delivering a wealth of actionable data to stakeholders allowing the streamlining and improvement of agricultural practices [18
The AquaCrop model is another decision-support tool used for modeling and devising strategies for the efficient management of crop-water productivity [19
]. This model is also capable of predicting crop productivity, water requirements and water-use efficiency under water-limiting conditions [20
]. Moreover, AquaCrop estimated the biological yield of rice more accurately compared to the Crop Environment Resource Synthesis-Rice and ORYZA2000 models [21
]. Other previous studies have used the AquaCrop model to simulate crop growth and yield in different environments [22
], ranging from potato [23
], soybean [24
], maize [9
], wheat [6
] to rice [30
]. Other studies, however, suggest that AquaCrop be further tested in different locations in order to increase the accuracy of crop production prediction in various conditions [19
In addition to crop models, remote sensing (RS) is another appropriate tool used to assess crop growth and production at various scales [33
]. RS data with coarse and middle resolution are widely used in rice cultivation research [35
]. Recently, several RS methods have been developed to map rice areas in different parts of the world, such as optical RS-based mapping, microwave RS-based mapping and the integration of the two methods [7
]. Using different techniques, optical images from such satellites as the Advanced Very High-Resolution Radiometer (AVHRR) with 1.1 km spatial resolution, the Moderate Resolution Imaging Spectroradiometer (MODIS) with 250, 500, 1000 m spatial resolution and the Landsat-Vegetation data set (30 m spatial resolution) have been used to estimate cropping area of rice and crop yield [20
]. Although these techniques demonstrate advantages for rice monitoring at regional to global scales because of their wide coverage and relatively long data archiving periods, the low resolution of satellite data are not suitable for rice crop mapping in fields that are relatively small in size, irregular in shape and sometimes fragmented by well-built roads and dense water networks, or mostly mixed with other land cover types. These small-scale rice fields are the critical limitation of RS methods because the area of some rice fields is smaller than the spatial resolution of these satellite platforms. As a result, the mixed-pixel problem is prominent, resulting in creating uncertainty in the discrimination of the spectral signatures of rice from other land cover type [40
]. Few studies have attempted to estimate the rice yield using high-resolution RS data (such as Quickbird; 0.65 m, WorldView; 0.31 m and IKONOS; NIR 3.2 m, PAN 0.82 m) in the past two decades [36
] but their approach has encountered problems with swath width and high cost [42
]. Using Landsat imagery has also encountered a problem with obtaining the cloud-free images due to the temporal resolution, as the revisit frequency is 16 days [7
]. As a result, it makes it impossible to obtain phenology information during the relevant crop period. Previous studies suggest the combination use of satellite data with an acceptable spatial resolution and a more frequent revisit cycle [40
] but again, it would result in high costs.
Despite shortfalls, a recent study found that imagery from the Chinese HJ-1A/B satellite (30 m spatial resolution and 4 days revisiting) has been successfully applied in China to rice-field mapping along with other crops such as rice, maize, sunflower, and wheat [40
]. Here in this study, we integrated moderate-resolution optical images from HJ1A/B (30 m) with the AquaCrop model to simulate the yield of transplanted rice on small farms in a Rice Seed Production Community as a pilot case study area to support agricultural planning policies in Thailand. In addition to using Aquacrop model, our study attempted to employ the advantages of HJ1A/B with an adequate spatial resolution for small farm (not less than 900 m2
or 1 pixel in the image) and high temporal observation.
RS data contain the growing condition of the crop at the observation time only, making it difficult to estimate crop yield directly [72
]. In addition, although RS-based empirical forecasting models are relatively simple to build, these models cannot take into account temporal changes in crop yields without long-term field experiments [7
]. Recently, the integration of RS data into crop growth models has been investigated as a way to estimate crop yield, becoming increasingly recognized as a potential tool for yield forecasting [7
]. Moreover, various types of satellite imagery need to be considered to fit the intended utilization and resolution (e.g., spatial, spectral and temporal). However, the limited synergy between RS-based methods and crop models is complicated and so many input parameters are required, including several biophysical parameters (e.g., soil and meteorological variables) and plant parameters (e.g., biomass, LAI and height, age, etc.) which are usually expensive, labor-intensive and time-consuming to acquire [6
]. In this study, we combined the advantages of the AquaCrop model [20
] and moderate-resolution HJ1A/B data to eliminate such limitations, achieving successful results with satisfactory validation.
4.1. Canopy Cover
The canopy cover (CC), detected as LAI from the RS data, is one of the most important variables for crop models [60
] but is hard to determine in the field. According to field experiment results (Figure 4
), there is a good relationship between the measured average LAI and rice age for the two experimental rice groups. This explains why the LAI value increases with rice age/growth stage. The leaf area of rice plants increases with growth, reaches a maximum around heading and decreases thereafter due to the death of lower leaves [73
]. The presence of this relationship shows that the field data collection was highly accurate for crop parameter calibration.
When LAI was converted to CC using Equation (3), the calibrated best K value for both experimental rice plots was 0.70, signifying that this is the optimal constant value to use when applying the LAI to the CC conversion formula for similar local rice cultivars. The maximum CC was set as a crop parameter along with other cultivar-specific parameters collected from the field (Table 2
). The model was calibrated with this set of parameters, producing satisfactory simulated yields (Section 3.1.3
). Consequently, the HJ1A/B images were selected to transform values to CC, using the regression equations (Figure 7
) and polynomial formulae with higher R2
and graph relationships closer to crop growth phenology [74
]; therefore, these were proposed for the surface reflectance NDVI to CC conversion model. With the previously calibrated parameters (35 pixels in 15 plots), the CC values from HJ1A/B data were substituted and the yield was simulated in 126 test pixels in 25 plots for validation (Section 3.2
). Furthermore, the maximum CC (CCx
) percentage was more constant in the plots transplanted by machine. The validated rice plots (group A) showed only slightly different CCx
values (average CCx
79.06%, standard deviation 0.06), while the manually transplanted rice plots (group B) had fluctuated CCx
; the average CCx
was 92.16% and the average standard deviation was 0.5. The LAI and CC were calculated using the surface reflectance NDVI of the HJ-1A/B images, producing significant crop phenology that could be applied in future crop models as well as the image classification for rice cultivation methods.
4.2. Yield Simulation
After the application of calibrated crop parameters for yield simulation, there was a strong relationship between the observed and simulated yields in Group B, which used manual transplanting. The Nash-Sutcliffe model efficiency coefficient (EF) was 0.97, indicating a near-perfect match between the observed and simulated yields. Willmott’s index of agreement or compatibility (d) is close to 1, which signifies an almost-perfect agreement between the observed and simulated yields. The RMSE of rice group B was 0.36, indicating a few errors. The NRMSE is 7.24%, indicating an excellent relationship between the measured and predicted yields. The simulated yield of Group A was evaluated as an acceptable result with statistical indicators of 0.40%, 0.60%, 1.35% and 18.99% for EF, d, RMSE and NRMSE, respectively. Nevertheless, the validation of the simulated yields presented good results (MAE = 0.159, RMSE = 0.182 t ha−1
= 0.88) as shown in Table 8
and Figure 8
In this study, we presented a new approach for rice yield estimation at the farm level by integration of moderate-resolution optical satellite imagery and a crop model. The AquaCrop model was proposed as one of the best models for rice yield prediction as it requires a lesser number of input parameters in comparison to other crop models. The RS data provided useful crop information for parameter calibration and canopy cover (CC) calculation. One benefit of this approach is the reduction in cost of field experiments/observations because the moderate-resolution imagery provides enough quality for crop information on small rice farms. The results show that the performance of the calibrated AquaCrop model was satisfactory compared to observed crop parameters from the small paddy areas, adequately simulating rice yield in the rainy season for individual farms with different planting dates, planting methods and rice varieties. Although small numbers of cultivar-specific crop parameters were applied for calibration, the simulated yield results were good.
The outcome set of the cultivar-specific parameters which were calibrated in this research could be applicable for initial model calibration in other paddy areas. Although the range of the constant value can be scaled down for model calibration before application, it would be necessary to adjust the user-specific parameters based on different regions and conditions. Furthermore, we found that CC calculation using surface reflectance NDVI from HJ-1A/B images, HJ1A/B can deliver useful information for the yield forecasting model and achieved satisfactory results in this case study of rice grown using the transplant method.
We conclude that assimilation of moderate-resolution HJ1A/B imagery can be efficiently used with the AquaCrop model to extract crop information that would otherwise be difficult and costly to collect by field observations. This method can, therefore, be suggested to farmers or field managers for use as part of community-based policy and other agriculture planning projects. We further suggest that this model can be implemented in other paddy areas. To increase even more accuracy of our study, future research should focus on the application of rice crop cycles to the dry season in continuous crop-cycle years, crop parameters (LAI, CC, agricultural practices, etc.) for more cultivar rice varieties and the number of test plots/area in order to improve the simulated yield accuracy and expand applicable areas. Concerning a different kind of the method of the rice plantation would be challenged in the future study. With regards to the pre-processing of HJ1A/B satellite data, we also suggest that the importance of spatial autocorrelation should be taken into account for improved classification accuracy in future studies. The future study could evaluate the optimize of various vegetation index for LAI/CC extraction that would be developed and integrated with radar imagery. Also, cloud eliminate techniques should improve the results on multi-aspect.
In addition, to increase the potential of yield forecasting and implementation, long-term meteorological data should be sufficiently developed in rice growing areas and countries for reliable estimates of rice production.