MONITORING DIGESTATE APPLICATION ON AGRICULTURAL CROPS USING SENTINEL-2 SATELLITE IMAGERY ^††thanks: © 2025 IEEE. Published in the 2025 IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2025), scheduled for 3 - 8 August 2025 in Brisbane, Australia. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE. Contact: Manager, Copyrights and Permissions / IEEE Service Center / 445 Hoes Lane / P.O. Box 1331 / Piscataway, NJ 08855-1331, USA. Telephone: + Intl. 908-562-3966. ^††thanks: This version is the accepted manuscript submitted to arXiv. The final version will be published in the Proceedings of IGARSS 2025 and available via IEEE Xplore. For citation, please refer to the published version in IGARSS 2025.

For the development of the experiments, Land Parcel Identification System (LPIS) data for 2023, including crop type and parcel geometry, were provided by the Greek Payment Authority ”OPEKEPE” for the EU’s Common Agricultural Policy (CAP) subsidies. These agricultural parcels are located in the region of Thessaly, Greece. Information on digestate application events, location, date and quantity of application was supplied by ”EPILEKTOS” SELECTED TEXTILE INDUSTRIAL COMPANY S.A. The initial dataset comprised 272 parcels, of which 97 underwent treatment, while the remaining parcels served as controls. The general approach, as illustrated in Fig. 1, involves analyzing the temporal changes in the spectral behavior of partially vegetated soils treated with EOM. This analysis was conducted using Sentinel-2 imagery. To ensure the temporal alignment of digestate application and satellite data, Sentinel-2 images were selected based on their proximity to the dates of application. Sentinel-2 data were extracted for each parcel. The Scene Classification Map (SCL) generated by Sen2Cor [17] atmospheric correction was filtered to include only values corresponding to vegetation, non-vegetated areas, and water. Additionally, a cloud coverage threshold of 20% was applied.

In total, seven spectral indices were computed. $EOMI_{1}$, $EOMI_{2}$, $EOMI_{3}$, $EOMI_{4}$ and $NBR_{2}$ (Eq. 1-5) were previously proposed to monitor digestate application. $EOMI_{3}$ (Eq.3) combines $EOMI_{1}$ and $EOMI_{2}$ and uses the red, NIR and SWIR bands of Sentinel-2 most impacted by solid EOM. In order to consider winter wheat vegetation, $EVI$ and $NDVI$ have also been used. Finally, the ratios between EOM and vegetation indices were calculated to account for the combined effects of growing vegetation and EOM application.

where $B02$, $B04$, $B08$, $B8A$, $B11$ and $B12$ refer to the Sentinel-2 spectral bands.

Spectral indices were computed across all parcels in the dataset for all available Sentinel-2 imagery, extended 30 days before and 30 days after the application of digestate, since in several cases, the implementation was carried out in multiple stages over a short time frame.

The main goal of this study is to use ML to identify whether an agricultural parcel has received digestate treatment. ”EPILEKTOS S.A.”’ data annotates whether a parcel has undergone treatment or not, as it contains information regarding the treatment date(s) and the quantity of EOM per field. We labeled the data with 0 for untreated fields and 1 for treated fields. The ML models were then trained to predict the correct label based on these events (Binary Classification problem).

Following feature engineering and the application of one-hot encoding to categorical labels (crop type), the dataset was randomly split into training and test sets using stratified sampling in an 80/20 ratio. Three ML models were selected for training: Random Forest, Gradient Boosting, and k-Nearest Neighbors. The hyperparameters for the models were selected as follows: for the Random Forest, 100 estimators were used with no maximum depth and the square root of the number of features as the maximum number of features. For k-Nearest Neighbors, 5 neighbors were chosen with uniform weights. For Gradient Boosting, 100 estimators were utilized with a learning rate of 0.1. Furthermore, a feed-forward neural network was constructed that consists of an input layer that receives the preprocessed features, followed by a single hidden layer with 128 neurons and a ReLU activation function to introduce non-linearity. The output layer uses a Softmax activation function to produce class probabilities. This architecture is trained using Cross-Entropy Loss and optimized with the Adam optimizer, which adapts the learning rate during training. The network is trained over 10 epochs. For all models, evaluation was conducted using a 5-fold cross-validation strategy based on Precision, Recall, and F1 Score statistics.

After constructing the dataset, the results of photo interpretation within the RGB scope, $NDVI$ and $EOMI_{2}$ reveal distinct insights into the field conditions. We first visualized a parcel where the application of digestate is detectable through changes in the $EOMI_{2}$ index, and compared this to its RGB image and $NDVI$ values (see Fig. 2). This particular parcel had several dates of digestate application, with the most recent being on 28-08-2023, and a significant amount of EOM was applied. Notably, we observed visible changes even in the RGB image. However, an interesting finding was that the $NDVI$ index showed minimal, if any, changes following the digestate application, in contrast to the $EOMI_{2}$, which clearly detected the changes. After conducting the photo interpretation for all parcels, we found that 48 out of 97 fields did not exhibit a specific change in the $EOMI_{2}$ index. Therefore, the recall of detecting the change solely through photo interpretation is 50.51%. Additionally, we collected the crop codes for the agricultural parcels to determine whether there is any relationship between the crop type and the appearance of visible change in the photo interpretation. Since many crop codes belong to the same broader family, we grouped the crops into four main categories: Cereals, Cotton, Industrial Crops, and Leguminous Crops. Fig. 3 shows how annotations of actual changes in the $EOMI_{2}$ index are distributed across these categories and seasons. The plot reveals that the application of digestate is more visible for some crop types than others. However, in most cases, the distribution is nearly equal. As a result, no definitive conclusion can be drawn from visual inspection of spectral indices regarding the relationship between crop type and digestate application. The same can be observed in the second plot, where the comparison is implemented per season.

We implemented various ML models to assess the effectiveness of Sentinel-2 imagery in monitoring EOM application, with a focus on digestate implementation in agricultural fields of different crop types, and the results are diplayed in Table 1. The Random Forest model excels in identifying the negative class (class 0), showing high recall and a decent F1-score for this class. However, its ability to correctly identify the positive class (class 1) is limited, with a significant drop in recall. This suggests that the model may be overly biased toward the majority class (class 0). A similar observation can be made for k-Nearest Neighbors and Gradient Boosting, although they score slightly lower(significantly lower in class 1 precision), with k-Nearest Neighbors showing a more notable decrease in performance. The performance of the Feed-Forward Neural Network is between Gradient Boosting and Random Forest, having the best score in recall.

A key contribution of this study is its focus on digestate practices in non-controlled, real-world settings, which have been under explored in remote sensing. The results demonstrate that Sentinel-2 imagery can detect digestate applications on agricultural parcels of various types, offering a scalable approach for monitoring agricultural practices. This result might enable landscape-level assessments that surpass traditional methods like farmer surveys. Challenges were noted, particularly the fact that the application of EOM is not always observable. Further research is needed to refine spectral indices and address variables like crop growth stages, seasonal effects, and weather conditions. ML models also showed promise, with Random Forest outperforming Gradient Boosting and k-NN in precision and F1 scores for EOM detection. The Neural Network achieved the highest score in recall. In further iterations of this research, we plan to increase the size of the training dataset through the addition of more ground truth, as it will help in integrating better ML models for operational digestate application using free, open-acess Sentinel-2 imagery. Finally, on the soil health spectra, studies show that overall microbial diversity remained stable, but species richness increased with digestate and mixed fertilization. The introduction of digestate also led to the emergence of new microbial taxa associated with nutrient cycling and carbon storage, suggesting potential benefits for long-term soil health and productivity [18].