University of Twente Student Theses


Deep learning for image time-series analysis: Application to crop mapping

Sood, Megha (2021) Deep learning for image time-series analysis: Application to crop mapping.

[img] PDF
Abstract:In the present times, with the dynamically changing climatic conditions, it becomes very important to monitor and estimate the acreage and production of crops to meet the requirements of the bulging population. Conventionally, crop classification was carried out using unsupervised and supervised methods which demanded a lot of manpower and time investment. This research explores the automation capability of deep learning models for finding hidden patterns in the dataset. Keeping this in mind, an optimized number of dates were selected for generating training data for the deep learning model. 1-Dimensional Convolutional Neural Networks (CNN) and its integration with Long Short Term Memory (CNN-LSTM) models were explored and optimized to produce classified outputs. The Modified Possibilistic c-Means (MPCM) Classifier was utilized to test its performance in terms of handling intra-class variability within a crop class and compared with the deep learning model outputs. Also, the dual-sensor approach was studied to find any additional information useful for classification. The study area chosen was Ludhiana district in the agricultural state of Punjab, India. The research was divided into two phases with maize (Zea Mays), mentha (Mint), and guava (Psidium guajava) as target crops in phase-1 while early potato (Solanum Tuberosum), mid potato, and late potato as target crops in phase-2. Time series data was used to resolve the issue of overlapping spectral signatures of crops. The freely available Sentinel-2 dataset integrated with the Landsat-8 dataset was used in phase-1 of the study while the Sentinel-2 dataset was utilized in the phase-2 study. Class-Based Sensor Independent Normalized Differential Vegetation Index (CBSI-NDVI) was used to enhance the target crop and separate it from the other non-interest crops in the study area. The optimum number of dates was selected for each target crop by performing separability analysis. The date combinations that gave maximum separation between the target crop class and other (non-interest) crop classes were selected to be optimum. Then the dataset generated was used to create training data for the deep learning models. Two deep learning models were explored and optimized which were the 1D-CNN model and an integrated CNN-LSTM model. The deep learning models handle the heterogeneity within a crop class (intra-class variability) very well which is still a challenge for a fuzzy classifier. Here, a fuzzy MPCM classifier was studied against the training data to test its capability of handling the heterogeneity within a crop class during classification and was compared with the outputs of deep learning models. Also, a dual-sensor approach was explored where the Sentinel-1 Synthetic Aperture Radar (SAR) dataset was used alongside optical data to test for any additional information extracted from the data. It was observed that the CBSI-NDVI index proved to be reliable in enhancing a specific crop in the study area. The optimum dates selected using separability analysis helped to reduce the temporal domain. In phase-1, six-date combinations came out to be optimum for Maize and Mentha while seven-date combinations came out to be optimum for Guava. In phase-2, four-date combinations came out to be optimum for early and mid-potato while three-date combinations came out to be optimum for late potato. A simpler CNN model architecture proved to be efficient for specific crop classification which consisted of two 1D-CNN layers followed by a max pool layer, a dropout layer, and a dense layer. The integrated CNN-LSTM model consisted of three pairs of 1D-CNN and LSTM layers, three dropout layers, two max pool layers, and a dense layer. It was observed that the hybrid model resulted in better classification outputs for the target crops from both phases. It was also observed that the hybrid CNN-LSTM model handled the intra-class variability (heterogeneity within a crop class) best when compared to 1D-CNN and MPCM classifiers. However, the MPCM classifier proved to handle heterogeneity within a crop class better than the CNN model. In the end, it was also observed that using a single sensor dataset yielded better classification results when compared to the dual-sensor dataset. After analyzing the results, the hybrid CNN-LSTM model provided the best classification results. Also, the fuzzy MPCM classifier performed similarly to a deep learning model in terms of handling heterogeneity within a crop class.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page