Deep learning fusion for dynamic cropland monitoring: integrating MUNet with Siamese networks and convolutional LSTM2D

Agricultural countries like India need strong change detection systems for land covers which directly or indirectly affects the cropland. Detecting changes presents significant challenges due to factors like availability of data, image quality, and optical image resolution, particularly in areas like cropland where there’s high intra-class variation. The features in cropland, including agricultural land, grassland, and vegetation areas, often yield similar results, complicating the identification process. To address this, a novel approach using domain adaptive feature extraction is proposed to enhance feature extraction. With a high-resolution satellite imagery dataset obtained from the GeoFen-2 satellite comprising 600 HRS images, the emphasis shifts towards feature extraction, where three innovative architectural designs are introduced. Firstly, a Siamese-MUNet model is proposed. Secondly, conv-LSTM2D with 3D max pooling is introduced to generate highly weighted feature maps. Lastly, conv-LSTM2D integrated with Siamese-MUNet is presented. All three models utilize the proposed Modified UNet (MUNet) as the base model for better feature extraction and to find hidden patterns in bi-temporal image pairs. The proposed models achieve significant results, boasting accuracies of 93.8%, 94.9%, and 94.2%, respectively. Furthermore, these models undergo rigorous testing using ROC analysis, covering 99% of the test area. Though all models are performing well conv-LSTM2D with 3D max pooling is outperforming by detecting minute changes at a pixel level. Beyond accuracy, precision, and f1-score performance metrics such as dice coefficient, and mean IoU, which are pivotal in segmentation tasks, are evaluated.