The demand for new tools for mass remote sensing of crops, combined with the open and free availability of satellite imagery, has prompted the development of new methods for crop classification. Because this classification is frequently required to be completed within a specific time frame, performance is also essential. In this work, we propose a new method that creates synthetic images by extracting satellite data at the pixel level, processing all available bands, as well as their data distributed over time considering images from multiple dates. With this approach, data from images of Sentinel-2 are used by a deep convolutional network system, which will extract the necessary information to discern between different types of crops over a year after being trained with data from previous years. Following the proposed methodology, it is possible to classify crops and distinguish between several crop classes while also being computationally low-cost. A software system that implements this method has been used in an area of Extremadura (Spain) as a complementary monitoring tool for the subsidies supported by the Common Agricultural Policy of the European Union.

The early and automatic identification of crops declared by farmers is essential for streamlining European Union Common Agricultural Policy (CAP) payment processes. Currently, field inspections are partial, expensive and entail a considerable delay in the process. Chronological satellite images of cultivated plots can be used so that neural networks can form the model of the declared crop. Once the patterns of a crop are obtained, the correspondence of the declaration with the model of the neural network can be systematically predicted, and can be used for monitoring the CAP. In this article, we propose a learning model with neural networks, using as examples of training the pixels of the cultivated plots from the satellite images over a period of time. We also propose using several years in the training model to generalise the patterns without linking them to the climatic characteristics of a specific year. The article also describes the use of the model in learning the multi-year pattern of tobacco cultivation with very good results.

Satellite crop identification processes are increasingly being used on a large scale, both to verify the crop and to improve production. As it is necessary to study phenological data over a period of time across a large territory, a lot of storage space is needed to save the satellite images and a lot of calculation time to analyse all this information. Sensing periods are usually established based on subjective expert criteria or previous experience. However, this decision may cause several differences when discriminating crop patterns, besides not guaranteeing good precision. These processes would greatly improve if the appropriate time periods could be found systematically using the minimum number of satellite images in the shortest possible time. In this paper, we propose a new methodology to determine a suitable sensing period for crop identification using Sentinel-2 images, applying hill climbing algorithms to the training sets of neural network models. We have used the method successfully in the 2020 Common Agricultural Policy campaign in the Extremadura region, Spain. The article also describes the use of the method in a case on tobacco detection in this region.

Today, machine learning applied to remote sensing data is used for crop detection. This makes it possible to not only monitor crops but also to detect pests, a lack of irrigation, or other problems. For systems that require high accuracy in crop identification, a large amount of data is required to generate reliable models. The more plots of and data on crop evolution used over time, the more reliable the models. Here, a study has been carried out to analyse neural network models trained with the Sentinel satellite’s 12 bands, compared to models that only use the NDVI, in order to choose the most suitable model in terms of the amount of storage, calculation time, accuracy, and precision. This study achieved a training time gain of 59.35% for NDVI models compared with 12-band models; however, models based on 12-band values are 1.96% more accurate than those trained with the NDVI alone when it comes to making predictions. The findings of this study could be of great interest to administrations, businesses, land managers, and researchers who use satellite image data mining techniques and wish to design an efficient system, particularly one with limited storage capacity and response times.