*4.4. Evaluation of Spring Wheat Yield at the Regional Scale*

Using a cropland mask, the calibrated model was applied in the study area. The spring wheat yield was evaluated based on the best model 4 from 2000 to 2017 in the Northern part of Mongolia (Selenge and Darkhan provinces) and the results are shown in (Figure 9). The temporal and spatial spring wheat yield maps produced based on the best model using NDWI (June), VSDI (third decade of June), and NDVI (first decade of July) indices, at the regional level have been generated in cropland area from 2000 to 2017. The spatial patterns of estimated spring wheat yield ranged from 1.3–35(100kg ha<sup>−</sup>1). In generated spring wheat yield maps, green colors indicating the normal and favorable condition and highest values of crop yield; red color shows unfavorable conditions and less amount of crop yield, respectively. From this result, we can see that spring wheat yield was low in drought years (2000–2002, 2005, 2006, 2015, and 2017). The irrigation system is limited in Mongolia, most of the spring crops at their vegetative and reproductive stages suffer water stress due to recurrent drought. Drought stress influences the water supply to vegetation and reduces accumulated biomass and production of crops [65].

**Figure 9.** Estimated temporal and spatial spring wheat yield map at the regional level.

#### **5. Discussions**

The study paves a new way for crop monitoring in Northern Mongolia. We have explored nine remote sensing indices in decade and month intervals. We found that there are 4 indices (VHI, VSDI, NDWI, and NDVI) that are more relevant than other indices for spring wheat yield estimation. This study has found that the NDWI and VSDI are the best indices for Mongolian crop monitoring. The NDWI was mainly indicated as an effective tool for water stress, soil, and vegetation moisture conditions and water content in vegetative areas, which was determined by the NIR and SWIR bands [11,12,33,48,62]. The supply of moisture to the north-central cropping region of Mongolia comes out as the main factor that clearly demonstrates the results of findings [66]. Water deficiency causes a physiological disorder that can inhibit cell division and differentiation, leading to the reduction of plant size and yield [67]. Our best results were obtained through Model 4 that showed R2 = 0.55, respectively. The results of the relationship between indices showed that MSAVI obtained the best wheat yield estimation model (R2 = 0.63), which was slightly higher than our result. Indices (SAVI, MSAVI, NDVI, and EVI) selected for wheat yield estimation in irrigated Indus Basin of Pakistan [68] are the difference from ours, because at, irrigation area, water is not stress issue. Water is a stressed factor in rainfed spring wheat in Mongolia. Dempewolf et al. [49] developed the wheat yield forecasting model in Punjab province of Pakistan using time-series MODIS and Landsat derived vegetation indices (NDVI, WDRVI, EVI2, SANDVI). The final results show that a forecasted wheat yield was within 0.2% and 11.5% of actual values, which was lower than our result. Bolton and Friedl et al. [62] compared the accuracy of different indices (EVI2, NDVI, and NDWI) in different zones for maize and soybean yield in the Central United States, and the EVI2 obtained best accuracy result with R2 = 0.73, respectively. This result was higher than our result.

We also find the highest correlation between indices with spring wheat yield and peaked at the flowering stage. The peak period of vegetative for spring wheat yield is June and July in this region. This implies that vegetation has its strongest response to moisture availability during this period. The growth condition of spring wheat during flowering stage might have more yield information than other stages at all growth stages, which means the developed regression model can be predicted two months before harvesting the crop and the correlation of estimated and actual yield from heading to flowering periods is higher than other crop growth stages. These results are in agreement with previous studies showing this to be the most suitable time to predict yield [27,48,60,61,69].

Furthermore, we find that later June is the most critical time for spring wheat yield formation. Indices in later June are used in every equation. We examined the relationship between actual spring wheat yield with 10 days and monthly indices for the regression model in the growing period (June–August). Similarly, we tried out the relationship between accumulated and relative values of nine indices with spring wheat yield from 2000 to 2017. The results of the correlation between accumulated and relative values of remote sensing nine indices with actual yield were lower than 10 day and monthly indices value. Juan Sui et al. [11] developed the dry aboveground mass and wheat yield estimation model using several remote sensing indices derived from MODIS and Himawari-8 sensor. A dry aboveground mass and yield errors of <10% and 12% were reported in Hengshui city of Hebei province, which was slightly lower than our results. Lopresti et al. [27] performed based on time-series MODIS-NDVI data for wheat yield estimation model obtained a higher correlation with estimated and actual yield (R<sup>2</sup> = 0.75), which was higher than our results. Also, several regression models for crop yield estimation based on MODIS, NDVI, are presented [18,48,61]. Moriondo et al. [61] have carried out the NDVI data to estimate wheat yield in the Grosseto and Foggia provinces of Italy. The results of the correlation coefficient between simulated and actual yield were 0.73–0.77, with corresponding RMSE were 0.44Mg/ha and 0.47Mg/ha, respectively.

The impacts of global warming are already confronted in Mongolia, visible from records between 1940 and 2013 from 48 meteorological stations. According to Dagvadorj et al. [70], the temperature has increased by 2.07◦C compare to mean. Due to the impacts of climate change, more extreme and continued droughts have occurred in many parts of Mongolia and which directly affect the vegetation and crop growth, biodiversity and socioeconomics in Mongolia [9]. Reportedly, [10,44,71] 2000–2002, 2004, 2005, 2007, and 2009 years were extremely affected by mild to severe drought and slight drought-hit Mongolia in 2003 and 2011. An additional notable finding of this study is that the spatial regional spring wheat yield distributions shown that the spring wheat yield was high in 2011, 2012, 2013, and 2016 and was low from 2000 to 2003 and 2015. It was statistically significant (*p* < 0.001), respectively and confirmed our result that during the drought years' spring wheat yield was low. Perhaps, in the year 2003, we had the highest precipitation in our monitoring period. The amount of precipitation, soil type, soil moisture, and changes in air temperature have a significant impact on wheat yield. Particularly, drought and soil moisture deficit influence the most reduced crop yield and vegetation size [72]. Thus, from our results, we recommend developing an irrigation system for spring wheat cultivation and increase the number of crop yield observation samples in this region. These results obviously show the promising application of NDWI and VSDI data in crop yield assessment at relatively cheap cost and timely.

## **6. Conclusions**

In Mongolia, the application of remote sensing methodology in agricultural policy and practices is in its nascent stage. This was the first time a multi-regression model based on remote sensing indicators was used to estimate crop yield in Northern Mongolia which is the main spring wheat-producing region. For this purpose, the best and most suitable indices were first defined through the testing of correlations between the nine indices and the actual spring wheat yield. Our results show that NDVI, NDWI, VCI, TCI, VHI, and NMDI indices with spring wheat yield were positively correlated (0.47, 0.51, 0.38, 0.47, 0.45, and 0.15), respectively and NDDI, VSWI, and VSDI with spring wheat yield were negatively correlated (−0.39, −0.57, and −0.38), respectively. Furthermore, the results confirmed the importance of the integration of both satellite and ground data for crop yield estimation. Consequently, we selected the NDWI, VSDI, and NDVI as the most suitable indices out of the nine indices, which are NDVI, NDWI, VCI, TCI, VHI, and NMDI. The highest negatively and positively correlated indices are a combination of NIR, red, blue, and SWIR bands. SWIR and red bands are found more sensitive to moisture variation and water stress of crops and soils [33]. Among nine indices, NDWI (0.51) in June and VSDI (−0.57) in July show the highest correlated indices with actual spring wheat yield, which indicates that the soil and crop moisture, as well as water content, are very important factors for crop yield.

As next step, we refined and developed the regression model using the above three selected indices in order to estimate crop yield. In total, six models were elaborated to be used for the growing period which is June to August. Timeline observation showed a higher correlation between indices and spring wheat yield during the flowering stage in June and July. Therefore, it is suggested that a suitable time to estimate spring wheat yield is at this stage of one to two months before the harvest. Whereas the results for the month of August showed a lower correlation indicating the lateness to estimate. The best results were obtained through Model 4 that used a combination of indicators from the period of third 10 days of June of VSDI, average of June NDWI and first 10 days of July of NDVI. Therefore, the Model 4 is the most effective predictor for crop yield monitoring in the northern part of Mongolia.

In this paper, we could estimate only 74% of the actual yield. This was due to several reasons that possibly could be ascribed to the different spatial resolution between MODIS data (1 km) and the ground measured spring wheat yield data (station-based measurements). Also, phenomena such as the different soil structures and the amount of precipitation have a big influence on the yield. However, the application of remote sensing regression model results enormously enrich the ground station collected data by providing large scale, region-wide data for the decision-makers to better manage food security challenges. In the future, work needs to be carried out to apply more consistently high-resolution images, such as Landsat and Sentinel for more accurate estimation of crop yield. In general, a comprehensive and systematic use of remote sensing technology in the agriculture sector of

Mongolia is to be considered, including broader policy for research and development, the introduction of the latest technology and equipment and targeted capacity-building activities.

**Author Contributions:** B.T. contributed to the research experiments, methodology, software, data analysis, and the writing-original draft of the manuscript; B.W. conceived the experiments, and was responsible for the research analysis; H.Z. contributed to experimental design and the manuscript revision; G.B. contributed to the collection of crop yield data and the manuscript revision; L.N. contributed the manuscript revision.

**Funding:** This paper was funded by the National Natural Science Foundation of China (41561144013, 41601464), National Key R&D Program of China (2016YFA0600304), the International Partnership Program of Chinese Academy of Sciences (131C11KYSB20160061,121311KYSB20170004), the Queensland-Chinese Academy of Sciences (Q-CAS) Collaborative Science Fund (grant number: 131211KYSB20170008 & 2017000257).

**Acknowledgments:** We are very grateful to colleagues of the Information and Research Institute of Meteorology, Hydrology, and Environment (IRIMHE) of Mongolia. Special thanks to Elbegjargal Nasanbat for his comments and suggestions. We thank the anonymous reviewers for reviewing the manuscript and providing comments to improve the manuscript.

**Conflicts of Interest:** The authors declare no conflict of interest.
