Abstract:
Forest and wildlife preservation departments in Pakistan face many challenges and are ill-equipped to counter them. Departmental incompetence, ambiguous resource rights, and judicial delays were examples of institutional weaknesses that impede forest monitoring. A novel approach has been devised through remote sensing and machine learning techniques to tackle temporal forest monitoring on a large scale. This study's specific objective was to develop an automatic system to classify and detect forest cover and its trend over 30 years over the Mansehra region, Khyber Pakhtunkhwa (KPK). Machine Learning (Random Forest) algorithm was used to classify and detect forest cover using Free-to-use satellite imageries (Landsat) and datasets. Segmentation was applied to labeled six land-cover classes (2015) and trained (2100 patches) and test (900 patches) classifier (Random Forest (RF)) with an accuracy of 89.47 % and 87.58 %, respectively. Random Forest classifier clearly outperforms the other statistical machine learning methods and is used for the complete analysis of classification (land-cover/ land-use) and forest cover change detection over the Mansehra Region. Training accuracy percentage was calculated for Random Forest classifier, Support vector machine (SVM), Classification and Regression Tree (CART), and Perceptron as 89.47, 66.33, 73.24, and 49.95, respectively. The testing accuracy percentage calculated for RF was 87.58, SVM was 46.81, CART was 78.48, and perceptron was 51.71. The forest cover area detected is 1495.6898 Km2, 1279.4715 Km2, 1129.5112 Km2, and 1461.0676 Km2 for 1900, 2000, 2010, and 2019 respectively. Overall forest cover change (1990-2019) was calculated using 'change detection metrics' with forest gain percentage (9.73%), loss percentage (10.56%), and percentage effective forest change (-0.83%). In few areas (Biari, Batang, and Behri), percentage effective change maybe a surprise and very effective, which means the afforestation in some areas generate overall forest gain percentage in past decades. Areas like Nokot, Kiwai, and Ghanian region's forest effective change percentage were calculated as -15.38, -17.01, and -2.36, respectively. The study's extension might help to get more cross-temporal data so that a more reliable system can be made for continuous monitoring of forest change.
