As in what type of method you would have to use, anything related to Machine Learning and Artificial Intelligence can usually be reduced to three aspects:

(1) Data recollection

(2) Model building and Validation

(3) Model interpretation and insights

Without having much knowledge about what particular crop or problem you are studying, you will have to gather a comprehensive database with experimental data measuring the variable you want to predict (for example total crop recolection, time to maturity, time to harvest...etc) and with measurements or estimations of the variables you expect to be important for your problem (in this case for example, water availability, temperature and CO2 concentration will probably be useful descriptors).

Once you have assembled a good dataset, you use ML to select the best attributes and then to build predictive models on your dataset.

Then you have to read and understand the predictive models, trying to learn insights about your problem from them.

Artificial intelligence can be used by measuring changes in temperature, air humidity, lack of soil moisture, and lack of watering of plants, which affects the accuracy and efficiency of predictions.

I used machine learning models for modeling missing data on actual sunshine duration, net radiation on the earth's surface, and evapotranspiration. I tested various approaches, models, and feature selection methods to determine how they affect modeling results. Many countries face challenges with the availability of certain types of measured data, and climatological stations are not available in every locality. In my country, we have enough climatological stations, but only some of them measure actual sunshine duration or radiation. At stations where these measurements are available, the time series often contain missing dates.

The value of actual sunshine duration (or net radiation) is important for calculating reference evapotranspiration (ET0) using the FAO56 P-M methodology. This can also be used (after multiplying by the crop coefficient) as an alternative to actual evapotranspiration. These variables are necessary inputs for many models used in the assessment of aridity, drought, the impact of climate change, and hydrological modeling.

I modeled these variables using other measured climatological data from the stations, remote sensing data, and freely available climatological data. Additionally, as a new approach to data homogenization, I used time series of the same variables (ETc, ET0) from similar stations nearby. The results showed that ML models provide very good predictions in all approaches, with precision significantly better than that of traditionally used models. I see great potential in using machine learning models for all types of tasks related to data modeling for assessing climate change.