Taking a mixed methods approach, this study draws on both quantitative and qualitative methods, namely land-cover change mapping and ethnographic interviews. More specifically, we take an “interactive design” to mixed methods, in that our approach emphasizes quantitative and qualitative methods equally. We do not, however, aim for data transformation (e.g., we do not take our qualitative data and transform them into quantifiable data for statistical analysis). Instead, we aim for “data importation”, namely that data from one approach was reflected upon mid-stream in our analysis, and fed into the analysis of the other data set, and
vice versa in an iterative manner [
69]. We began by mapping land-cover change via remote sensing. We then chose cases representing the most important dynamics between LULCC and livelihoods, and drew on previous interviews completed since 1999 regarding livelihoods as well as additional interviews and observations focusing on the case sites during the summers of 2012, 2013, and 2014 to guide our interpretations. An important element in this interactive process was our long-term knowledge of the region, which allowed us to be confident about our choice of cases.
4.1. LULC and LULCC Mapping
Given the lack of accurate data regarding land use in rural Vietnam, especially historical data [
70], we had to compromise somewhat methodologically, opting for an analysis of a mixed system of land-use and land-cover types. LULC maps were derived from Landsat thematic mapper (TM5) and enhanced thematic mapper (ETM+) images obtained from the United States Geological Survey (USGS). All images were taken during the winter season (
Table 1) when sparse vegetable crops were growing, with no maize or wet rice. To identify land-cover types from our Landsat images, we drew on the Level I land-cover types proposed by Anderson in 1976 [
71] and previous LULCC studies of Vietnam using Landsat images [
52,
61].
Table 1.
Dates of images.
Table 1.
Dates of images.
Province | Dates |
---|
Lai Châu | 2009 November 03 |
2000 November 02 |
Lào Cai | 2009 November 12 |
1999 December 27 |
Hà Giang | 2009 November 05 |
2000 November 04 |
Initially, we identified five key land-
cover types across the study region from satellite images (shrubs, bare soil, open canopy trees, closed canopy trees, and water). To provide further detail, we then added one land-
use type, namely built-up areas, working from the five original land-
cover types. It should be noted that there is rarely any grassland in this mountainous area; when small patches of grass are mixed up with shrubs, we identify them only as “
shrubs”. Cropped land is composed of subsistence crops (lowland and upland rice, corn, cassava) and cash crops (banana, pineapples, tobacco, among others). According to our interviews and local crop calendars, areas with subsistence crops are often covered by
bare soil in the winter, when our images were acquired. This is also observed in other upland areas of Vietnam [
4,
52,
61]. Pineapples are small and fairly dispersed plants, hence pineapple areas are spectrally similar to
bare soil in the images. Areas having bananas remain vegetated and identified as
shrubs in the Landsat images. However, bananas were not very common in the region in 2009 (interviews and observations), and were usually found next to houses or mixed with shrubs. Built-up areas are also identified as
bare soil in the images, although dense urban areas have stronger spectral reflectance. To separate built-up from cropped land, we used information regarding road systems (see below).
Open canopy trees in this region tend to be bamboo, plantations for timber production, or natural succession since the 1990s (interviews).
Closed canopy trees are old and nearly intact forests that are often protected by the local government, such as special-use forests (often labeled national parks). This class has sometimes been called “mature and evergreen forest” in Southeast Asian forest studies [
72]. Separating forests into two categories—open and closed canopy—is usual in Southeast Asian upland studies [
4], although some authors tend to group them into one category [
61,
73]. Interpretation of classification results was done with these details in mind. Illustrations of these LULC types are presented in
Figure 2, while we detail our classification method below.
The images were pre-processed to remove distortions caused by sensor errors, atmospheric interference, and surface irregularities. Cloud and cloud shadow masking was conducted on the 2000 image of Lai Châu and the 2009 image of Lào Cai (more specific technical details are provided in [
14]). We used an object-based approach [
74] to identify the five land-cover types listed above plus clouds and shadow. Although associated most often with very high-resolution images, this approach has proven accurate to produce rural land-cover types from middle-resolution images such as Landsat [
75,
76]. Cultivation systems in upland Southeast Asia are complex, mixing young forest, shrubs, and different types of crops [
52,
72]. In this context, an object-based approach is very helpful in creating segments (objects) and incorporating textural information of segments into classification. This approach has also allowed us to work at two spatial levels: segmentation of large-size patches (to separate large-scale forest and crops) and of small-size patches with relatively homogenous texture and signals (to identify smaller plots).
Figure 2.
Illustration of the six LULC classes. (Photo credit: Lê Mạnh An, Thi-Thanh-Hiên Pham).
Figure 2.
Illustration of the six LULC classes. (Photo credit: Lê Mạnh An, Thi-Thanh-Hiên Pham).
Segmentation parameters in eCognition (bands, scales, color/shape ratio, and compactness/smoothness ratio) were tested at different values. Segmentations and rule-based classifications were undertaken at different scales using the same band composition (bands 1–5, 7), color/shape ratio (0.2/0.8), and compactness/smoothness ratio (0.3/0.7). Segmentation values of 5, 10, 20, 50, 100, and 200 were tested, aiming at creating segments of different sizes. Segments were visually examined to determine the visibility of the main land-cover classes. We chose two values that produced the most homogenous segments in terms of spectral values and texture. The first segmentation was conducted at a scale of 50 to obtain large size objects. Rules were then used to classify those objects into clouds and shadow, water, and bare soil classes. The second segmentation was conducted at a scale of 10, then a second set of rules was used to classify segments into three vegetation classes of shrubs, open canopy trees, and closed canopy trees.
Bare soil was then separated into bare soil and built-up. Built-up pixels were assigned by evaluating road density. The road network was separated into solid (gravel or concrete) and non-solid (compacted soil susceptible to flooding during the rainy season) roads. A density map was developed by assigning solid roads an importance value of 3 and non-solid roads a value of 1. Pixels having high road density (3m/km2 in 1999 and 4m/km2 in 2009) were visually compared to aerial images on Google Earth and those that coincided with more urban areas (built-up and having a street network) were renamed as built-up. The remaining bare soil pixels remained bare soil.
We did not use a set of training points per se to define rules of classification. Rather, we chose roughly 200 objects (segments) that corresponded to land-cover types that we knew on the ground from observations. Then we created rules composed of textural and spectral indicators for each class from the chosen objects. Refining and adjusting rules were based on interactive “trial and error”. The classification process included initial field observations in summer 2013, and field verification in summer 2014 by the second author.
4.2. Assessing the LULC Mapping
A ground truth assessment was conducted for the 2009 classification by using ground control points (GCPs). There are no historical air photos for the region covering 1999/2000. There are a few photos on Google Earth from 2009/2010 but they are dispersed and cover only 15 percent of the study area. We hence opted for GCPs. We collected 365 GCPs: 142 points in Lào Cai Province (September 2012), 101 points in Lai Châu, and 122 points in Hà Giang (September 2013). Given the difficult access to many locations in the region, points were sampled along roads, focusing on typical land-cover types in the region. There were fewer points in Lai Châu given the lack of roads and access difficulties caused by landslides.
Each point was registered in GPS (precision of 5m) and photographed in the four cardinal directions to capture potential mixtures of land cover. Descriptions of land use and land cover were included at each point and for the four directions. Land-cover types were assigned to each point based on photos and descriptions. Unfortunately, we do not have ground points for the 1999 images, but since we used the same image processing procedure for the 1999 and 2009 images, we believe that the accuracies of the 1999 images are similar to those of the 2009 images.
In order to evaluate the accuracy of the 2009 classification, we created a confusion matrix (
Table 2) using the 365 GCPs. The overall accuracy is 71 percent, mostly due to confusions between
open canopy and
closed canopy classes. Accuracies varied from 46 percent (
open canopy) to 96.88 percent (
closed canopy). In this mountainous area undergoing complex forest transitions,
open canopy trees are usually mixed with
shrubs, making it difficult to separate them from Landsat images. The confusion between
open canopy and
closed canopy is most likely due to the fact that open forests in the three provinces are highly heterogeneous. What is defined on the ground as
open canopy in areas with plantations (since the mid-1990s) could be similar spectrally to
closed canopy in other areas. When grouping these two categories into “forest” (as done in several Southeast Asian LULCC studies [
61,
73]), we obtained a “user accuracy” of 79.30 percent and a “producer accuracy” of 72.22 percent, raising overall accuracy to 73.70 percent.
Other important confusions related to
bare soil, our proxy for rice or corn fields. In these mountainous areas, dispersed houses and remote roads are often located close to rice or corn fields, creating a spectral mixture in the image. In Đồng Văn District, Hà Giang Province, this gets complicated further by limestone outcrops within fields and near houses (
Figure 3a). Lastly, the confusion between
shrubs and
bare soil is explained by the fact that upland dry rice and maize are commonly planted in areas close to shrubs (
Figure 3b).
Table 2.
Confusion matrix and producer’s and user’s accuracy for accuracy assessment of the 2009 classification (based on ground control points).
Table 2.
Confusion matrix and producer’s and user’s accuracy for accuracy assessment of the 2009 classification (based on ground control points).
| Ground Reference | Total | Producer Acc. (%) |
---|
Water | Closed Canopy | Open Canopy | Shrubs | Bare Soil | Built-Up |
---|
Classification | Water | 16 | 0 | 0 | 0 | 1 | 2 | 19 | 84.21 |
Closed canopy | 3 | 31 | 10 | 7 | 3 | 0 | 54 | 57.41 |
Open canopy | 0 | 1 | 23 | 8 | 2 | 2 | 36 | 63.89 |
Shrubs | 0 | 0 | 13 | 42 | 7 | 2 | 64 | 65.63 |
Bare soil | 3 | 0 | 3 | 10 | 109 | 11 | 136 | 80.15 |
Built-up | 0 | 0 | 1 | 3 | 15 | 37 | 56 | 66.07 |
Total | 22 | 32 | 50 | 70 | 137 | 54 | 365 | |
User Acc. (%) | 72.73 | 96.88 | 46.00 | 60.00 | 79.56 | 68.52 | | |
Figure 3.
(a) Typical karst landscape and (b) mixture of maize and shrubs in Đồng Văn District, Hà Giang Province. (Photo credit: Sarah Turner and Lê Mạnh An).
Figure 3.
(a) Typical karst landscape and (b) mixture of maize and shrubs in Đồng Văn District, Hà Giang Province. (Photo credit: Sarah Turner and Lê Mạnh An).
4.4. Qualitative Fieldwork
The qualitative fieldwork that supports this study was completed by the first author during repeated research visits to Lào Cai Province since 1999, Lai Châu Province since 2004, and Hà Giang Province since 2009, and by the second author in Lào Cai Province since 2012. This includes over 100 in-depth unstructured (conversational) interviews with ethnic minority farmers (Tày, Hmong, Yao, Nùng) in Lào Cai Province, and 50 each in Lai Châu and Hà Giang Provinces. Approximately 75 percent of farmer interviewees were women, as they were more likely to be in the house during the day, yet had a wide knowledge of land uses and changes. Farmer ages ranged from 25 to 80 years old. Interviews ranged in duration from 20 minutes to over an hour, focusing on local livelihood diversification, agricultural practices, land-use changes, and state-society relations. Twenty-eight semi-structured interviews were also completed with provincial officials (mostly Kinh) working in government departments linked to agriculture, planning, labor, and natural resources. The core themes of these interviews were livelihoods, market integration, land-cover change, and the impacts of state policies on each of these. Interviews were completed with the aid of local ethnic minority interpreters for farmers of the same ethnicity, or with Kinh interpreters or alone for Kinh farmers, urban dwellers, and state officials. All interviews were transcribed and coded using a mix of constant comparative, axial, and thematic qualitative coding approaches. Concurrently, observations of LULCC have been completed and noted annually.