Development of Deep Learning Algorithm for Remote Sensing Image Classifications to Assess the Impacts of Coal Mining on Land Use/Land Cover Patterns

Abstract

Accurate classification of the satellite image is always a challenging task, and hence uncertainty involves in the change detection analysis in the land use/land cover (LULC) for mining regions. The present study attempts to demonstrate the development of an optimized deep convolutional neural network (DCNN) using Linear imaging self-scanning sensor-IV (LISS-IV) data for the classification of satellites image to assess the LULC pattern in a mining region. The primary requirement of the DCNN model development is an image dataset for different LULC types for training and validation of the model. In this study, an image dataset was initially prepared from the LISS-IV satellite image of the study region for optimized model development, and thereafter the model was tested with the Landsat dataset. The common LULC types identified in the specified mining regions are Barren Land (BL), Built-Up Area (BA), Coal Mining Region (CM), Vegetation (VE), and Waterbody (WB), respectively. The false-color composite (FCC) of three bands [B2 (Green): 0.52 - 0.59 μm, B3 (Red): 0.62 - 0.68 μm, and B4 (Near-infrared): 0.77- 0.86 μm] was used for extracting the image database. The image databases were prepared for three different benchmarks or dimensions (6×6, 12×12, and 24×24) for examining the effect of input image size on model performances. The number of image samples derived for each LULC type is 1250. Thus, a total of 6250 image samples of five classes (1250 for each class) were derived and labelled with class numbers and types. Out of 6250 LISS-IV image datasets, 70% of the data was assigned for training, and the rest 30% was assigned for validation of the DCNN model. The study examined the performances of the model using three optimizers including the Adaptive Moment Estimation (ADM), Root Mean Square Propagation (RMSPro), and Stochastic Gradient Descent with Momentum (SGDM) for identifying the best one by fixing the values of hyperparameters. Moreover, the values of the learning rate, momentum, minibatch size, maximum epoch, learning rate drop factor, learning rate drop period, L2Regularization, verbose frequency, and validation frequency were fixed as 0.001, 0.1, 10, 0.95, 0.0001, 300, 128, 50, and 30, respectively in the model for identifying the best optimizer. The performances of the DCNN model were found to be 99.04%, 97.1%, and 97.5% respectively with SGDM, RMSProp, and ADAM. The results indicate that the SGDM optimizer offers the highest accuracy as compared to others with the specified model hyperparameters and thus SGDM optimizer was further used for the optimization of the model hyperparameters through sensitivity analysis. The DCNN model hyperparameters were tuned in terms of the learning rate, epoch number, batch size, and momentum to obtain the best output. The model offers satisfactory results in terms of accuracy for both the training dataset (=99.04%) and the validation dataset (=87.50%). The classification accuracy was further evaluated on the testing dataset by randomly choosing the samples through visualization of the Google Earth Image Pro of the same time frame of the study area. Additionally, the model performances were measured using the six indices (accuracy, error, precision, recall, F1-score, and MCC), which were derived from the confusion matrix parameters (true positive, false positive, true negative, and false negative). After successful testing of the model with the LISS IV dataset, the performance of the same was further evaluated using Landsat images. For this, a new image database of (6 × 6) size was prepared from Landsat image for evaluating the optimized DCNN model. In this case, the total number of image samples extracted was the same as that used for LISS IV data. The results indicate that the DCNN model offers better accuracy with the Landsat data (training dataset: 94.73% and validation dataset: 89.03%). Moreover, the long-term data of the Landsat sensor are freely available and thus can be used for time-series analysis. Detection and delineation of coal mining regions using remote sensing data is a challenging task as the characteristics of the surface features are very close to barren lands. Thus, the study also examined the efficacy of the optimized DCNN model in the detection and delineation of coal mining regions (Jharia Coalfield). The study examined the effect of the image size of the training database on model performances. The model performances were tested using three different image size databases [DB6 ∈ (6×6), DB12 ∈ (12×12), and DB24 ∈ (24×24)]. The results indicated that the classification accuracies with DB6, DB12, and DB24 training and validation datasets are nearly the same (>99%) in each case but the boundary delineation with lower size image training dataset was more smooth. Therefore, the dataset of DB6 ∈ (6×6) was used for further study. The third objective of the study is to make a comparative evaluation of the DCNN and DNN-based LULC classification of mining regions using multi-sensors (LISS-IV, Landsat-8, and Sentinel-2A) fused data. The study designed DCNN and deep neural network (DNN) algorithms for LULC classification of the multi-sensor fused image, respectively. A discrete cosine transform (DCT) with a spatial correlation technique was used to derive the fused data from three different satellite sensor data. LULC for the specified region was classified into the same five broad categories including, barren land, built-up area, coal mining region, vegetation, and waterbody. The results reveal that the DCNN model consistently outperforms the DNN model, showcasing accuracy, error rates, precision, and recall ranging from 99.83% to 99.99%, 0.01% to 0.17%, 99.52% to 99.99%, and 99.40% to 99.99% on the training dataset, and 99.50% to 99.99%, 0.01% to 0.50%, 98.35% to 99.99%, and 98.33% to 99.99% on the validation dataset, respectively. In comparison, the DNN model demonstrates values ranging from 90.36% to 99.90%, 0.01% to 9.64%, 75.10% to 99.53%, and 66.99% to 99.99% on the training dataset, and 88.50% to 99.94%, 0.06% to 11.50%, 72.25% to 99.66%, and 62.50% to 99.99% on the validation dataset. These findings showed that the DCNN classification algorithm outperforms the DNN classification algorithm. Moreover, the comparative performances of the DCNN model with different datasets indicate that the model with fused images outperformed the model with individual sensor images. The last objective of the study is to make a time-series analysis of satellite data through transfer learning for assessing the impacts of coal mining activities on LULC change. The long-term impacts of mining activities in Jharia coalfield (JCF) on LULC patterns using transfer learning of the DCNN model. The study used three bands (Band 3, Band 5, Band 7 of Landsat 5 and Landsat 7, and Band 4, Band 6, Band 7 of Landsat 8) of Landsat series data from 1987 to 2021 at an interval of two years for time-series analysis. A new image database with a large number of image samples was prepared from the Landsat series data for generating the base model using optimized model hyperparameters. A total of 2000 image samples of 6×6 size were prepared for each class for base model development to classify the LULC into five different classes (barren land, built-up area, coal mining region, vegetation, and waterbody), and thus the total number of image samples was 10000. The image database was partitioned into training and validation of the proposed DCNN model in the ratio of 7: 3. The study results revealed that the model offers an accuracy level of 95% and 88 % on the training and the validation dataset, respectively. The base model learning algorithm was subsequently transferred for classifying the time-series Landsat data to evaluate the long-term impacts of mining activities on land-use patterns. The results indicate that barren land, coal mining region, and waterbody have decreased from 237.30 sq. km. (=39.88 %) to 171.25 sq. km (=28.78 %), 118.77 sq. km. (=19.96 %) to 68.73 sq. km (=11.55 %), and 35.58 sq. km (=5.98 %) to 18.68 sq. km (=3.14 %) during 1987 to 2021, respectively. On the other hand, the built-up area and vegetation have increased from 120.14 sq. km (=20.19 %) to 233.02 sq. km (=39.16 %) and 83.19 sq. km (=13.98 %) to 103.36 sq. km (=17.37 %) during 1987 to 2021. The study also analyzed the time-series correlation to understand the sensitivity of transforming one land-use type into other. The time-series correlation results indicate that coal mining is the most sensitive land use type from 1987 to 2021, whereas barren land is least sensitive up to 2011, and thereafter vegetation is the least sensitive LULC class.