Automatization of an early cereal maps classification model from 2010/2011 to 2021/2022 in a semi-arid region

Cereals represent an essential factor in the economy and food security, especially for countries that are not self-sufficient and depend on imports like Tunisia. Obtaining an early cereal map without the need to collect field data and without waiting for the end of the agriculture season helps the government to make early decisions. 

Hence, the first objective of our study is the development of an automatic classification model first calibrated for one agricultural year, 2020/2021 (2021), and then validated over the years 2011 through 2022 in the Kairouan governorate. The second objective is the development of a forecasting model in order to have early cereal maps several months before the harvest which occurs in June. 

Using Sentinel 2 and Landsat 5-7-8 data, different vegetation indices percentiles have been calculated. In order to select the best indices for cereal classification, a feature importance study over all the indexes was performed using the random forest classification algorithm reference year classification. A land cover classification model was validated for the reference year 2021, with an overall accuracy of 89.3%. This classifier has been used to elaborate land cover classification maps since 2011, focusing mainly on cereal crops. Using Sentinel 2 data, a good precision (P) for cereal crops was found, between 85,8% and 95,1%. Good to moderate accuracies were obtained when using Landsat data, between 41% and 91,8%. Then, a land cover forecasting model was validated for 11 years for different forecasting periods where we found excellent results four months before harvest (in February). We were able to obtain the cereal crop maps with a P between 85,1% and 95,1% using Sentinel 2 data and between 42,6% and 95,4% using Landsat data from four months before harvest. However, confusion between cereals and cereals grown with arboriculture was found which is due to the similarity between these two classes.

With this automatic land cover model, we have been able to produce the cereal maps of the last 12 agricultural years. This approach could be also used in the future to obtain a cereal map as early as February.