|Paper title||Understanding Neural Networks for Crop Yield Estimation|
|Form of presentation||Poster|
Improving the management of agricultural areas and crop production is strictly necessary in the advent of global population growth and the current climate emergency. Nowadays, several methodologies at different regional and continental scales exist for monitoring croplands and estimating yield. In all schemes, Earth observation (EO) satellite data offer massive, reliable and up-to-date information for monitoring crops and characterizing their status and health efficiently in near-real-time.
In this work, we explore and focus on the potential of neural networks (NN) for developing interpretable crop yield models. We ingest multi-source and multi-resolution time series of satellite and climatic data to develop the models. We focus on the interpretability in the case study of the larger area of the US Corn Belt. The study area is one of the leading agricultural productivity regions globally due to its massive production of cereals. Particularly, we have built models to estimate the yield of corn, soybean and wheat. According to previous studies, the synergy of variables from different sources has proven successful [1,2,3]. As input variables, we selected a variety of remote sensing and climatic products sensitive to crop, atmosphere, and soil conditions (e.g., enhanced vegetation index, temperature, or soil moisture). Neural networks provided excellent results in all crops (R>0.75) matching other standard regression methods like Gaussian processes and random forest.
Understanding neural networks is of utmost relevance, especially with overparameterized and neural networks. Interpreting what the models learned allows us to extract and discover new rules governing crop system dynamics, such as the influence of the input variables, rank agropractices, and study the impact of climate extremes (such as droughts and heatwaves) on production. And all these in a spatially explicit and temporally resolved manner. In addition, temporal data streams allow us to detect which temporal instant is more critical along the different phenological states of the crop regarding productivity terms. For this purpose, we explore several techniques to shed light on what the trained neural networks learned from EO and crop yield data such as methods to study the activation of different neurons on NN, and the associated with different time instants. These experiments open up new opportunities to understand crop systems and justify the necessary management decisions in order to enhance agricultural control in a changing climate.
 Mateo-Sanchis, A., Piles, M., Muñoz-Marí, J., Adsuara, J. E., Pérez-Suay, A., Camps-Valls, G. (2019). Synergistic integration of optical and microwave satellite data for crop yield estimation. Remote sensing of environment, 234, 111460.
 Martínez-Ferrer, L., Piles, M., Camps-Valls, G. (2020). Crop Yield Estimation and Interpretability With Gaussian Processes. IEEE Geoscience and Remote Sensing Letters.
 Mateo-Sanchis, A., Piles, M., Amorós-López, J., Muñoz-Marí, J., Adsuara, J. E., Moreno-Martínez, Á., Camps-Valls, G. (2021). Learning main drivers of crop progress and failure in Europe with interpretable machine learning. International Journal of Applied Earth Observation and Geoinformation, 104, 102574.