Using contours of objects from Sentinel satellite images for classifying agricultural lands using neural networks
Abstract
The study aims to enhance the accuracy of classifying agricultural lands as arable or non-arable using Sentinel-1 and Sentinel-2 satellite images by integrating spectral and structural indicators into deep neural networks as additional features. Proposed use of a specially developed structural indicator – the anthropogenic landscape degradation indicator (LDI), which quantitatively reflects the density and contrast of object contours on the terrain and serves as a marker of anthropogenic load. The aim of this work is to assess the effectiveness of LDI as an additional feature in deep learning of agrolandscape classification models and to compare its informativeness with the classic vegetation index, NDVI (Normalized Difference Vegetation Index). The research methodology involves the use of a neural network with the U-Net architecture with the EfficientNetV2-L encoder, pre-trained on the ImageNet-21k dataset, using attention mechanisms that aggregate temporal features in the process of processing multispectral and radar images. Experimental tests were conducted on a sample covering more than 90 thousand land plots in the Berlin-Brandenburg region (Germany), divided into arable and non-arable lands. The modeling results showed that the use of the LDI index as an additional feature increases the classification accuracy from 88.87% to 91.34% due to a more complete consideration of the structural features of the agricultural landscape. A comparative analysis with NDVI demonstrated the superiority of LDI in the tasks of distinguishing degraded and abandoned lands. Moreover, the highest accuracy rate (91.98%) was obtained with the joint use of NDVI and LDI, which indicates the effectiveness of combining spectral and structural indicators. The scientific novelty of the study lies in the introduction of the structural indicator LDI as a new parameter for neural network models of agricultural land classification
References
2. Adegun, A.A. et al. (2023) Review of Deep Learning Methods for Remote Sensing Satellite Image Classification. Journal of Big Data.
3. Aleissaee, A.A., Kumar, A., Anwer, R.M. et al. (2023) Transformers in Remote Sensing: A Survey. Remote Sens, 15(7), P. 1860.
4. Mehmood, A., et al. (2021) Vision Transformers for Remote Sensing Applications: A Review. Remote Sens, 13(3), P. 516.
5. Li, X., et al. (2022) Multi-Source Data Fusion for Land Cover Classification Using Sentinel-1 and Sentinel-2 Imagery. ISPRS J. Photogramm. Remote Sens, 183, P. 210–223.
