Optimized 1D-CNN model for medicinal Psyllium Husk crop mapping with temporal optical satellite data
2022
Sabir, Anam | Anil Kumar,
Deep Learning models are preferred for complex image analysis-based solutions to application-oriented problems. However, the architecture of such models largely influences the results which includes several hyperparameters that need to be tuned. This study aims at developing an optimized 1D-CNN model for medicinal Psyllium Husk crop mapping using open source temporal optical Sentinel-2A/2B satellite data. In this study, a sequential 1D-CNN model architecture was developed by optimizing hyperparameters which includes convolution layers, number of neurons, activation function, and batch size. Psyllium Husk crop fields were mapped in the Jalore district of Rajasthan using Sentinel 2A/ 2B (10 m) optical data. For spectral dimensionality reduction of the data, Modified Soil Adjusted Vegetation Index (MSAVI2) was used to maintain the data dimensionality since temporal data was utilized. The dataset was subsequently refined to include the target crop's specific phenological stages that distinguish it from other closely resembling species. The information corresponding to these specific crop stages was fed to the 1D-CNN model to carry out the classification. A range of training sample sizes were explored to determine the optimal number of training data points. As the output from the model, fractional images are obtained consisting of values proportional to the probability of a pixel lying in the target class. Accuracy assessment was carried out using fuzzy error matrix (FERM) by generating fractional output images from temporal optical PlanetScope data (3m) which was used as a reference. The best overall accuracy among the test cases came out to be 89.85% using conventional MSAVI2 with 1000 training samples.
显示更多 [+] 显示较少 [-]