Irrigation uses 70% of the world’s fresh water and large parts of the planet are already suffering water scarcity, from California’s Central Valley, the Mid-West, and including large portions of Asia, Southern Europe, and Australia. Due to the impact irrigation has on our planet, our team is interested in tackling the issue of agricultural irrigation detection and irrigation fraud detection. As we began researching the topic and the solutions that exist, we identified BigEarthNet as a promising dataset for our project. Unfortunately, we couldn't locate many articles that discussed the dataset and the effects the various features have on machine learning models. Our team decided to fill that void and perform an ablation study on the BigEarthNet dataset to help future researchers quickly identify which features have the most impact on machine learning models.
"BigEarthNet is a new large-scale Sentinel-2 benchmark archive, consisting of 590,326 Sentinel-2 image patches. To construct BigEarthNet, 125 Sentinel-2 tiles acquired between June 2017 and May 2018 over the 10 countries (Austria, Belgium, Finland, Ireland, Kosovo, Lithuania, Luxembourg, Portugal, Serbia, Switzerland) of Europe were initially selected. All the tiles were atmospherically corrected by the Sentinel-2 Level 2A product generation and formatting tool (sen2cor). Then, they were divided into 590,326 non-overlapping image patches. Each image patch was annotated by the multiple land-cover classes (i.e., multi-labels) that were provided from the CORINE Land Cover database of the year 2018 (CLC 2018)" - bigearth.net
BigEarthNet is 66 GB (compressed). Each folder in the dataset consists of an image tile. In each folder, there are GeoTiff files for each band of the image (R, G, B, NIR, etc.). The bands are at differing resolutions, so to stack the bands to recreate the original image, image modification methods must be used.
We used ResNet50 to identify agricultural land from BigEarthNet after converting the BigEarthNet labels to binary labels (Agricultural / Not Agricultural). Unfortunately, we ended up being more hardware restricted than we had anticipated. As such, we were only able to run 4 epochs per experiment, so there could still be noise in our results.
| Bands | Accuracy | Loss |
|---|---|---|
| B02, B03, B04 (RGB) | 0.7868 | 0.4210 |
| All Bands | 0.8198 | 0.3877 |
| B01, B02, B03, B04, B05, B06, B07 (RGB, Coastal Aerosol, Vegetation Red Edge) | 0.8260 | 0.3790 |
| B01, B05, B06, B07 (Coastal Aerosol, Vegetation Red Edge) | 0.8272 | 0.3751 |
| B05, B06, B07, B11, B12, B8A (Vegetation Red Edge, SWIR, Narrow NIR) | 0.8419 | 0.3481 |
Through our study of the BigEarthNet dataset, we found that the most influential bands, when predicting agricultural land, were the Vegetation Red Edge bands. Our model that performed the best on the test dataset was the model that was trained on Vegetation Red Edge(B05, B06, B07) and all the infrared bands (B08, B8A, B11, B12). Our initial hypothesis that the RGB channels would be the most important was proven false. We hope that our study of the BigEarthNet dataset will act as a foundation to help researchers better tackle illegal irrigation identification. Reducing illegal irrigation should help mitigate the damage irrigation has on our environment as well as give a larger water supply to the general populous, especially those currently without access to fresh water.