Land cover multiclass classification of wonosobo, Indonesia with time series-based one-dimensional deep learning model

Abstract

Mapping accurate land cover is critical to support authorities in producing a better land management policy, particularly in Wonosobo, which suffers from increased erosion and sedimentation due to extensive land conversion. Unfortunately, the conventional land cover map is mainly done by manual digitation on the screen, which is highly ineffective and time-consuming. This study presents a lightweight Deep Learning (DL) model to automate land cover mapping. We used one-dimensional CNN (Convolution Neuron Network) or CNN1D alongside Bi-GRU (Gated Recurrent Unit) and FCN (Fully Connected Network) layers to process pixel-based multivariate time-series dataset extracted from satellite imagery of Landsat 8 OLI (Operational Land Imager). Sampling points were gathered from the ground survey, and visual inspection via Google Earth was used to extract pixel values from multi-temporal imageries. During the model fitting which repeated 10 times, our model delivered a consistent performance with a mean peak validation score of 0.925 and a mean training score of 0.929. In the experiments stage, we added two more band indices into the main features; however, our model performance did not experience a noticeable improvement. Compared to other well-known DL models as well as traditional ML models, our model performed better and showed strong stability in accuracy, precision, and f1-score metrics for each land cover class. These results show that our proposed model can potentially improve the efficiency of the automation of land cover mapping.