**4. Conclusions**

By integrating satellite imagery and hydroclimatic information using a machine learning algorithm, we created annual irrigation maps for a subhumid area in southwestern Michigan for 2001–2016. The maps capture the spatiotemporal pattern of irrigation at a high spatial resolution (30 m) and indicate that irrigated area in southwestern Michigan roughly tripled in the last 16 years according to linear regression.

We demonstrated the utility of novel input variables including weather-sensitive remote sensing, spatial anomalies, and recently-developed composite indices. In particular, we found that those vegetation indices following dry periods are the most important to distinguish irrigated fields from rainfed. This not only reduces the number of scenes (thus memory and computational expense) to process, but also avoids possible confounding effects of high vegetation indices captured during a wet period.

The annual irrigation maps are validated using multiple data sources. Reasonable accuracy is achieved despite the difficulties involved with estimating irrigated area in a region with a subhumid climate and heterogeneous agricultural managemen<sup>t</sup> practices (e.g., deficit irrigation strategy for seed corn). We found that the mapping accuracy in dry years is higher than in wet years with a narrow margin. The small difference between accuracies may be attributed to the use of spatial anomaly and weather-sensitive remote sensing indices, which were able to distinguish irrigated from rainfed fields even under subhumid conditions.

We identified several challenges and limitations for mapping irrigated areas in subhumid to humid regions, including the dependency on the quality of input data (e.g., land cover) and cloud coverage, which is more frequent in such regions. The substantial efforts and difficulty involved in generating training data are also noteworthy and call for in season high-resolution imagery. Nevertheless, the promising results underscore the potential of using remote sensing and cloud computing to provide valuable information for water resources decision makers and hydrologic studies at regional scales.

**Supplementary Materials:** The annual irrigation maps 2001–2016 can be downloaded at https://doi.org/10. 4211/hs.3766845be72d45969fca21530a67bb2d. In addition, the following are available online at http://www. mdpi.com/2072-4292/11/3/370/s1, Table S1: The mean and quantiles for the cumulative probabilities 0.025 and 0.975 of number of available scenes for all pixels in the study domain between June 10th and August 5th for each year in the study period (2001–2016), Table S2: All input variables of the random forest classifier grouped into seven categories, Table S3. Unsuccessful input variables that were not used in the final random forest classifier.

**Author Contributions:** T.X., J.M.D., A.D.K., and D.W.H. conceived and planned the experiments. T.X. carried out the experiments. T.X. and J.M.D. developed codes. T.X., J.M.D., and A.D.K contributed to preparation of data. T.X. drafted the manuscript. All authors contributed to analytical methods, interpretation of results, and editing of the manuscript.

**Funding:** This work was supported by USDA NIFA Grants 2015-68007-23133 and 2018-67003-27406, and NSF Grants 1027253 and 1637653. We thank Jeremy Rapp for his contribution to training dataset preparation. We are grateful for the thoughtful review and constructive comments by the four anonymous reviewers.

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