24.08.23 - ELECTS-LSTM: Early Crop Classification Using Satellite Imagery and AI
The ability to classify crops on the basis of satellite imagery would greatly assist the authorities in controlling agricultural subsidiaries.
Currently, most approaches focus on the accuracy of predictions and are based on data from the whole year. This requires a lot of storage and computing resources. In addition, early predictions are very important for time-sensitive applications. Consequently, the ECEO laboratory at EPFL (École polytechnique fédérale de Lausanne) developed an End-to-End Learned Early Classification of Time Series (ELECTS) model for in-season crop type mapping.
They choose to use a Long Short-Term Memory (LSTM) recurrent neural network accompanied with a loss function incorporating both a cross-entropy loss and an earliness reward. With each new image in the temporal sequence, the fields are classified and if their confidence is high enough, the processed is stopped for the concerned elements.
The model was trained and tested on two large European datasets (BreizhCrops, BavarianCrops) and on two smaller African datasets (Ghana, South Sudan). This allowed to test its applicability to different regions of the world and to smaller datasets. The type of crops could vary from one region to another.
The model achieved an overall accuracy (OA) of 80% or more on the European datasets, with at most 40% of the entire sequence length. The OA is as high as that of state-of-the-art models working on complete time series.
In South Sudan, the OA reaches 83%, which is close to an other convolutional LSTM model, but 6% lower than what can be obtained by a random forest classified. In Ghana, the model achieves an OA of 54%, which is lower than the convolutional LSTM model and the random forest by 7.1% and 5.9% respectively. In both case, the average day of classification is in March, which allows to save a much smaller data volume compared to full series.
The ELECTS-LSTM model provides near state-of-the-art results without any region-specific parameter tuning. It is even on par with accuracy-only models on sufficiently large datasets, while classifying the crops earlier in the year.
The problem of crop classification was already discussed within the STDL but has not yet led to a project. Maybe we will have the occasion to test this methodology in the years to come.
#crop #agriculture #neuralnetwork